Files
nvidia__skills/benchmarks.json
Moshe Abramovitch 215153b017 chore(benchmarks): regenerate benchmarks.json for the new parser fields
The `check` job requires benchmarks.json to match what the aggregator would
generate, so a generator change and its output have to land together. Running
the official generator (not a hand edit) produces:

  - the three new v3 provenance fields on every entry
  - nemo-fabric-integrate and nemo-fabric-build-adapter

Those two skills were already missing on main: the regeneration in #466
merged before the sync in #465, so the committed index captured 341 skills
while skills/ held 343. `check` has been failing on main since, independent
of this branch. This regeneration also repairs that.

The new guard was exercised by this run and correctly did not fire — nothing
lost values, two skills gained them.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Moshe Abramovitch <moshea@nvidia.com>
2026-08-20 11:37:44 -05:00

35780 lines
887 KiB
JSON

{
"schema_version": 2,
"source": "skills/*/BENCHMARK.md",
"skill_count": 343,
"result_row_count": 3200,
"skills_without_results": [
"deepstream-generate-pipeline",
"digital-health-clinical-asr-setup",
"dynamo-interconnect-check",
"dynamo-recipe-runner",
"dynamo-router-starter",
"dynamo-troubleshoot",
"earth2studio-create-datasource",
"earth2studio-create-prognostic",
"earth2studio-deterministic-forecast",
"mcore-create-issue",
"mcore-linting-and-formatting",
"mcore-run-on-slurm",
"mcore-split-pr",
"mcore-testing",
"nemoclaw-user-guide",
"omniverse-realtime-viewer",
"physical-ai-infrastructure-setup-and-resilient-scaling",
"rag-blueprint",
"rag-eval",
"rag-perf"
],
"skills": [
{
"skill": "accelerated-computing-cudf",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 13,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"has_results": true,
"average_uplift_pct": 11.0
},
{
"skill": "aiq-deploy",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"has_results": true,
"average_uplift_pct": 1.6
},
{
"skill": "aiq-research",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "aiq-research",
"component": "AIQ",
"has_results": true,
"average_uplift_pct": 18.1
},
{
"skill": "amc-run-rtsp-calibration",
"evaluation_date": "2026-07-09",
"profile": "internal",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 20.6
},
{
"skill": "amc-run-sample-calibration",
"evaluation_date": "2026-07-09",
"profile": "internal",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 27.0
},
{
"skill": "amc-run-video-calibration",
"evaluation_date": "2026-07-09",
"profile": "internal",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 29.6
},
{
"skill": "amc-setup-calibration-stack",
"evaluation_date": "2026-07-09",
"profile": "internal",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 23.3
},
{
"skill": "cudaq-guide",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 9,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"has_results": true,
"average_uplift_pct": 11.7
},
{
"skill": "cuopt-developer",
"evaluation_date": "2026-06-26",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 23.7
},
{
"skill": "cuopt-install",
"evaluation_date": "2026-06-26",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 48.0
},
{
"skill": "cuopt-multi-objective-exploration",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:45bef5ee60d2a85ef8c4a7175c3d9a15eeac738e7bcf291d44de12ad4a35fff2",
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 27.8
},
{
"skill": "cuopt-numerical-optimization-api",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:f385c69ce235035e8e5cbfd91a8d3245b75473f0bf0d694451e988e439f6e9e1",
"validation_status": null,
"tasks": 9,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 35.2
},
{
"skill": "cuopt-numerical-optimization-formulation",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:7f39d48b2b7e50bd2a24f83c7f6172f1f16b1895a59e237bb4c345cf690cb71e",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 39.6
},
{
"skill": "cuopt-routing-api-python",
"evaluation_date": "2026-08-12",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:e13146ac43d578682f0a3d799837f67335d277ef722feb1d28fc526709b6ba4e",
"validation_status": null,
"tasks": 7,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 31.2
},
{
"skill": "cuopt-server-api-python",
"evaluation_date": "2026-06-29",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"has_results": true,
"average_uplift_pct": 50.5
},
{
"skill": "cupynumeric-hdf5",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 17,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"has_results": true,
"average_uplift_pct": 12.6
},
{
"skill": "cupynumeric-install",
"evaluation_date": "2026-07-27",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 24,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"has_results": true,
"average_uplift_pct": 31.4
},
{
"skill": "cupynumeric-migration-readiness",
"evaluation_date": "2026-07-27",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 27,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"has_results": true,
"average_uplift_pct": 26.2
},
{
"skill": "cupynumeric-parallel-data-load",
"evaluation_date": "2026-07-27",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 7,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"has_results": true,
"average_uplift_pct": 25.2
},
{
"skill": "dali-dynamic-mode",
"evaluation_date": "2026-08-11",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.0",
"dataset_digest": "sha256:cd54de141f8255b7043b6f2d8b5a203dad6038b1c20b0e36a549f4fac0721cba",
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"has_results": true,
"average_uplift_pct": 32.4
},
{
"skill": "data-designer",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "data-designer",
"component": "Data Designer",
"has_results": true,
"average_uplift_pct": 7.2
},
{
"skill": "deepstream-dev",
"evaluation_date": "2026-06-12",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 7,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 10.3
},
{
"skill": "skill",
"evaluation_date": "2026-06-15",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "deepstream-generate-pipeline",
"component": "DeepStream",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "deepstream-import-vision-model",
"evaluation_date": "2026-06-12",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 15.8
},
{
"skill": "deepstream-profile-pipeline",
"evaluation_date": "2026-06-03",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 5.8
},
{
"skill": "deepstream-run-mv3dt",
"evaluation_date": "2026-07-02",
"profile": "internal",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"codex"
],
"catalog_dir": "deepstream-run-mv3dt",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 13.2
},
{
"skill": "deepstream-sop",
"evaluation_date": "2026-06-25",
"profile": "internal",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 20.0
},
{
"skill": "dicom-metadata-extract",
"evaluation_date": "2026-05-31",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 12.3
},
{
"skill": "dicom-series-preflight",
"evaluation_date": "2026-05-31",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 5.9
},
{
"skill": "dicom-series-to-volume",
"evaluation_date": "2026-05-31",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 6.5
},
{
"skill": "digital-health-clinical-asr-build",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"has_results": true,
"average_uplift_pct": 12.1
},
{
"skill": "digital-health-clinical-asr-eval",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"has_results": true,
"average_uplift_pct": 16.0
},
{
"skill": "digital-health-clinical-asr-finetune",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"has_results": true,
"average_uplift_pct": 20.7
},
{
"skill": "digital-health-clinical-asr-setup",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "digital-health-clinical-asr-setup",
"component": "Digital Health",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "doca-aes-gcm",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 31.7
},
{
"skill": "doca-argp",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-argp",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 30.5
},
{
"skill": "doca-argus",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-argus",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 42.9
},
{
"skill": "doca-bare-metal-deployment",
"evaluation_date": "2026-07-25",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.6
},
{
"skill": "doca-bench",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-bench",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 33.9
},
{
"skill": "doca-bench-extension",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 36.6
},
{
"skill": "doca-bf3-deployment",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 37.6
},
{
"skill": "doca-bf4-deployment",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 29.7
},
{
"skill": "doca-caps",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-caps",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 35.9
},
{
"skill": "doca-collectx-deployment",
"evaluation_date": "2026-07-25",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 30.7
},
{
"skill": "doca-comch",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-comch",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 25.0
},
{
"skill": "doca-comm-channel-admin",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.9
},
{
"skill": "doca-common",
"evaluation_date": "2026-07-24",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-common",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 34.7
},
{
"skill": "doca-compress",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-compress",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 26.2
},
{
"skill": "doca-container-deployment",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.5
},
{
"skill": "doca-debug",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-debug",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 39.0
},
{
"skill": "doca-devemu",
"evaluation_date": "2026-07-25",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-devemu",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 35.9
},
{
"skill": "doca-dma",
"evaluation_date": "2026-07-24",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-dma",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 31.5
},
{
"skill": "doca-dms",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-dms",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 36.4
},
{
"skill": "doca-dpa",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-dpa",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 40.7
},
{
"skill": "doca-dpa-hl-tracer",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "doca-dpdk-bridge",
"evaluation_date": "2026-07-14",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 39.7
},
{
"skill": "doca-erasure-coding",
"evaluation_date": "2026-07-14",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 46.9
},
{
"skill": "doca-eth",
"evaluation_date": "2026-07-14",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-eth",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 44.3
},
{
"skill": "doca-firefly",
"evaluation_date": "2026-07-24",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-firefly",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 31.2
},
{
"skill": "doca-flow",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-flow",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 40.6
},
{
"skill": "doca-flow-dpa-perf",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 40.6
},
{
"skill": "doca-flow-dpa-provider",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 47.0
},
{
"skill": "doca-flow-grpc-server",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "doca-flow-perf",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 41.0
},
{
"skill": "doca-flow-tune",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 44.0
},
{
"skill": "doca-gpi",
"evaluation_date": "2026-07-24",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-gpi",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "doca-gpunetio",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 29.0
},
{
"skill": "doca-gpunetio-ib-write-bw",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 41.3
},
{
"skill": "doca-gpunetio-ib-write-lat",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.8
},
{
"skill": "doca-hardware-safety",
"evaluation_date": "2026-07-25",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 36.7
},
{
"skill": "doca-mgmt",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 31.1
},
{
"skill": "doca-pcc",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-pcc",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 35.0
},
{
"skill": "doca-pcc-counters",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "doca-pcc-ztr-rttcc-algo",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.5
},
{
"skill": "doca-programming-guide",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 32.5
},
{
"skill": "doca-public-knowledge-map",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.1
},
{
"skill": "doca-rdma",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-rdma",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 35.6
},
{
"skill": "doca-rdmi",
"evaluation_date": "2026-07-24",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 39.3
},
{
"skill": "doca-rmax",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-rmax",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 39.3
},
{
"skill": "doca-setup",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-setup",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 30.2
},
{
"skill": "doca-sha",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-sha",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 29.7
},
{
"skill": "doca-sha-offload-engine",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 34.2
},
{
"skill": "doca-socket-relay",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 33.1
},
{
"skill": "doca-spcx-cc",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 39.6
},
{
"skill": "doca-sta",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-sta",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 43.0
},
{
"skill": "doca-structured-tools-contract",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 32.6
},
{
"skill": "doca-telemetry",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 32.9
},
{
"skill": "doca-telemetry-exporter",
"evaluation_date": "2026-07-29",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.0",
"dataset_digest": "sha256:bdb053d3eebd3787bba933e197d64385efe318e5e5666e43c6300e91e2c8f42e",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 37.1
},
{
"skill": "doca-telemetry-utils",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 40.1
},
{
"skill": "doca-upgrade",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 23.6
},
{
"skill": "doca-urom",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-urom",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "doca-urom-svc",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 35.5
},
{
"skill": "doca-verbs",
"evaluation_date": "2026-07-26",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-verbs",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 35.1
},
{
"skill": "doca-version",
"evaluation_date": "2026-07-23",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "doca-version",
"component": "DOCA",
"has_results": true,
"average_uplift_pct": 41.5
},
{
"skill": "dynamo-interconnect-check",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "dynamo-interconnect-check",
"component": "Dynamo",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "dynamo-recipe-runner",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "dynamo-recipe-runner",
"component": "Dynamo",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "dynamo-router-starter",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "dynamo-router-starter",
"component": "Dynamo",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "dynamo-troubleshoot",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "dynamo-troubleshoot",
"component": "Dynamo",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "earth2studio-create-datasource",
"evaluation_date": "2026-06-01",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "earth2studio-create-datasource",
"component": "Earth2Studio",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "earth2studio-create-diagnostic",
"evaluation_date": "2026-06-24",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"has_results": true,
"average_uplift_pct": 10.3
},
{
"skill": "earth2studio-create-prognostic",
"evaluation_date": "2026-06-03",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "earth2studio-create-prognostic",
"component": "Earth2Studio",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "earth2studio-data-fetch",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "FAIL",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"has_results": true,
"average_uplift_pct": 1.6
},
{
"skill": "earth2studio-deterministic-forecast",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "earth2studio-deterministic-forecast",
"component": "Earth2Studio",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "earth2studio-discover",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"has_results": true,
"average_uplift_pct": 0.5
},
{
"skill": "earth2studio-install",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"has_results": true,
"average_uplift_pct": 1.1
},
{
"skill": "holohub-app-lifecycle",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:018f6fca61cefe2c611e9a1335321e75a24c166dbe6980ffbacfbc0f9d571733",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"has_results": true,
"average_uplift_pct": 23.1
},
{
"skill": "holohub-debug-build-run",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:c260676bf566fd39aec781cf254c16365214ab7b05e60c0d7f4ac468b89597dc",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"has_results": true,
"average_uplift_pct": 4.7
},
{
"skill": "holohub-module-lifecycle",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:cbdb11772e761dfe177d00b040909b44f6199ac0f2c1b37d2545a438328f2c52",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"has_results": true,
"average_uplift_pct": 16.3
},
{
"skill": "holoscan-install-conda",
"evaluation_date": "2026-05-31",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"has_results": true,
"average_uplift_pct": 6.2
},
{
"skill": "holoscan-install-container",
"evaluation_date": "2026-05-31",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"has_results": true,
"average_uplift_pct": 10.9
},
{
"skill": "holoscan-install-debian",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"has_results": true,
"average_uplift_pct": 10.8
},
{
"skill": "holoscan-install-source",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"has_results": true,
"average_uplift_pct": 13.7
},
{
"skill": "holoscan-install-wheel",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"has_results": true,
"average_uplift_pct": 5.8
},
{
"skill": "holoscan-setup",
"evaluation_date": "2026-06-29",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"has_results": true,
"average_uplift_pct": 22.8
},
{
"skill": "hsb-app",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 18.2
},
{
"skill": "hsb-flash",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "FAIL",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 18.2
},
{
"skill": "hsb-ip-create-top",
"evaluation_date": "2026-06-26",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 31.5
},
{
"skill": "hsb-ip-def",
"evaluation_date": "2026-06-26",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 34.4
},
{
"skill": "hsb-ip-packetizer",
"evaluation_date": "2026-06-26",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 47.5
},
{
"skill": "hsb-setup",
"evaluation_date": "2026-06-30",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 1.5
},
{
"skill": "hsb-test",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"has_results": true,
"average_uplift_pct": 18.4
},
{
"skill": "i4h-catheter-navigation",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 29.6
},
{
"skill": "i4h-catheter-navigation-digital-twin",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 25.0
},
{
"skill": "i4h-catheter-navigation-e2e",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 25.5
},
{
"skill": "i4h-catheter-navigation-render-drr",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 34.6
},
{
"skill": "i4h-catheter-navigation-setup",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 25.5
},
{
"skill": "i4h-catheter-navigation-smoke",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 21.0
},
{
"skill": "i4h-catheter-navigation-viewport",
"evaluation_date": "2026-07-20",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 27.5
},
{
"skill": "i4h-lerobot-viz",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 35.3
},
{
"skill": "i4h-workflow",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 26.3
},
{
"skill": "i4h-workflow-create",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 39.5
},
{
"skill": "i4h-workflow-dataset-annotate",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 33.5
},
{
"skill": "i4h-workflow-dataset-convert",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 25.5
},
{
"skill": "i4h-workflow-dataset-mimic",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 26.4
},
{
"skill": "i4h-workflow-dataset-replay",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 39.5
},
{
"skill": "i4h-workflow-dataset-teleop",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 37.8
},
{
"skill": "i4h-workflow-e2e",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 28.4
},
{
"skill": "i4h-workflow-finetune",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 37.4
},
{
"skill": "i4h-workflow-scene-edit",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 24.6
},
{
"skill": "i4h-workflow-setup",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 45.7
},
{
"skill": "i4h-workflow-validate",
"evaluation_date": "2026-07-03",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 9,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"has_results": true,
"average_uplift_pct": 32.0
},
{
"skill": "jetson-build-source",
"evaluation_date": "2026-05-30",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 15.2
},
{
"skill": "jetson-customize-camera",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 27.3
},
{
"skill": "jetson-customize-clocks",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 54.1
},
{
"skill": "jetson-customize-fan",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 35.1
},
{
"skill": "jetson-customize-mgbe",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 41.5
},
{
"skill": "jetson-customize-nvpmodel",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 39.6
},
{
"skill": "jetson-customize-pcie",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 28.3
},
{
"skill": "jetson-customize-pinmux",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 38.0
},
{
"skill": "jetson-customize-uphy",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 54.3
},
{
"skill": "jetson-customize-usb",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 25.7
},
{
"skill": "jetson-derive-carrier",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 33.7
},
{
"skill": "jetson-diagnostic",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 23.2
},
{
"skill": "jetson-download-bsp",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 24.6
},
{
"skill": "jetson-flash-image",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 39.8
},
{
"skill": "jetson-generate-kb",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 25.7
},
{
"skill": "jetson-headless-mode",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 25.1
},
{
"skill": "jetson-inference-mem-tune",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 34.1
},
{
"skill": "jetson-init-image",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 30.0
},
{
"skill": "jetson-init-source",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 23.9
},
{
"skill": "jetson-init-target",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 4.9
},
{
"skill": "jetson-link-docs",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 20.7
},
{
"skill": "jetson-llm-benchmark",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 32.0
},
{
"skill": "jetson-llm-serve",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 36.1
},
{
"skill": "jetson-memory-audit",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 18.2
},
{
"skill": "jetson-optimize-memory",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 39.1
},
{
"skill": "jetson-package",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 28.5
},
{
"skill": "jetson-print-bsp-info",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 22.8
},
{
"skill": "jetson-print-device-info",
"evaluation_date": "2026-06-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 22.1
},
{
"skill": "jetson-promote-image",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 38.0
},
{
"skill": "jetson-quick-start",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 32.9
},
{
"skill": "jetson-set-target",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 27.3
},
{
"skill": "jetson-speculative-decoding",
"evaluation_date": "2026-06-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 34.2
},
{
"skill": "jetson-validate-image",
"evaluation_date": "2026-06-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"has_results": true,
"average_uplift_pct": 11.0
},
{
"skill": "jetson-video-benchmark",
"evaluation_date": "2026-08-10",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.1.2",
"dataset_digest": "sha256:e2d80ff3bff2819a83593a6167c1fd9119caa0d8febe727d4a947e1b84baec88",
"validation_status": null,
"tasks": 7,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 46.4
},
{
"skill": "jetson-video-capability",
"evaluation_date": "2026-08-10",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.1.2",
"dataset_digest": "sha256:b440c2dac3de92def79952306c0425803168bede746d14213e367b04f0655282",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 24.5
},
{
"skill": "jetson-video-pipeline",
"evaluation_date": "2026-08-10",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.1.2",
"dataset_digest": "sha256:2954de192ec504a13b63de5d546a514eafdc11544903c203f84432c93bb7a5be",
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 43.0
},
{
"skill": "jetson-video-recipe",
"evaluation_date": "2026-08-10",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.1.2",
"dataset_digest": "sha256:cb2ef860b7b7cf59666c858c733f9b9fab262a8b7c52227977b0d53f6ae69da4",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 36.3
},
{
"skill": "jetson-video-setup",
"evaluation_date": "2026-08-10",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.1.2",
"dataset_digest": "sha256:da01efb8a143276d497b34e33588f31b5cb1d0f128bfa23256ee67a894ccd664",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"has_results": true,
"average_uplift_pct": 51.8
},
{
"skill": "launch-nemo-rl",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"has_results": true,
"average_uplift_pct": 12.9
},
{
"skill": "mcore-create-issue",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "mcore-create-issue",
"component": "Megatron-Core",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "mcore-linting-and-formatting",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "mcore-linting-and-formatting",
"component": "Megatron-Core",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "mcore-run-on-slurm",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "mcore-run-on-slurm",
"component": "Megatron-Core",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "mcore-split-pr",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "mcore-split-pr",
"component": "Megatron-Core",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "mcore-testing",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "mcore-testing",
"component": "Megatron-Core",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "nemo-automodel-distributed-training",
"evaluation_date": "2026-07-31",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:5a1d8d0db77adff7b1823958919deaac1b8ca472f0809af87c530be247ff8de0",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"has_results": true,
"average_uplift_pct": 32.7
},
{
"skill": "nemo-automodel-launcher-config",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"has_results": true,
"average_uplift_pct": 1.1
},
{
"skill": "nemo-automodel-model-onboarding",
"evaluation_date": "2026-07-31",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:814bfc7c94da8ea6fd1a065d7f3f0c4fcda9ef918f1da7eb46630700f241fc25",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"has_results": true,
"average_uplift_pct": 51.3
},
{
"skill": "nemo-automodel-recipe-development",
"evaluation_date": "2026-06-26",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"has_results": true,
"average_uplift_pct": 32.5
},
{
"skill": "nemo-fabric-build-adapter",
"evaluation_date": "2026-08-19",
"profile": null,
"environment": "local",
"evaluator_version": "1.3.2",
"dataset_digest": "sha256:b36cb8cb012011c6dbaaec7b2918d4dc18f18a2117ce8eab572181eeeb245976",
"validation_status": "passed",
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"has_results": true,
"average_uplift_pct": 25.9
},
{
"skill": "nemo-fabric-integrate",
"evaluation_date": "2026-08-19",
"profile": null,
"environment": "local",
"evaluator_version": "1.3.2",
"dataset_digest": "sha256:460ab15c5720276319ee246cc40eebe66a640b3874108e59b6e7c0e086991ebc",
"validation_status": "passed",
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"has_results": true,
"average_uplift_pct": 26.6
},
{
"skill": "nemo-mbridge-mlm-bridge-training",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 51.0
},
{
"skill": "nemo-mbridge-multi-node-slurm",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 1.7
},
{
"skill": "nemo-mbridge-perf-activation-recompute",
"evaluation_date": "2026-08-16",
"profile": null,
"environment": "local",
"evaluator_version": "1.2.7",
"dataset_digest": "sha256:b55bb3c51b83cb2245c0d7a8969bbac7066177e342bd9b9c316ac7cf26212ef6",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 36.5
},
{
"skill": "nemo-mbridge-perf-cpu-offloading",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": -0.5
},
{
"skill": "nemo-mbridge-perf-cuda-graphs",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 48.5
},
{
"skill": "nemo-mbridge-perf-expert-parallel-overlap",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:473eeb29e4297235c32a27b47541793b4f7eddba5f4fc3e053f65a5b344292f1",
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 47.3
},
{
"skill": "nemo-mbridge-perf-hierarchical-context-parallel",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 0.1
},
{
"skill": "nemo-mbridge-perf-megatron-fsdp",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 0.4
},
{
"skill": "nemo-mbridge-perf-memory-tuning",
"evaluation_date": "2026-06-30",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 59.7
},
{
"skill": "nemo-mbridge-perf-moe-comm-overlap",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:27e48e74e4c441a61624d9c376db3091241d8249209badb9c045fa6c966d02b8",
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 58.5
},
{
"skill": "nemo-mbridge-perf-moe-dispatcher-selection",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": -0.6
},
{
"skill": "nemo-mbridge-perf-moe-hardware-configs",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:bfc3173e4e2204740fad6a3746aa0b031118f0a404945c92ba4138ee08607b57",
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 48.0
},
{
"skill": "nemo-mbridge-perf-moe-long-context",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": -3.5
},
{
"skill": "nemo-mbridge-perf-moe-optimization-workflow",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:d195b5b0364709939e2d51e5857908465fae4f4bd91ec22a67656cf1acc798e9",
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 33.8
},
{
"skill": "nemo-mbridge-perf-moe-vlm-training",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 0.0
},
{
"skill": "nemo-mbridge-perf-parallelism-strategies",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 0.2
},
{
"skill": "nemo-mbridge-perf-sequence-packing",
"evaluation_date": "2026-08-11",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.0",
"dataset_digest": "sha256:57d3c088dee48ee97547a666ee24dcac45c6cd5cbf699640c3c7d41710f629a1",
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 36.3
},
{
"skill": "nemo-mbridge-perf-tp-dp-comm-overlap",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 0.5
},
{
"skill": "nemo-mbridge-recipe-recommender",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:ef1ce0be1100bdb45047e4131f641cc24d4f26a6a08e1a4f84d519d0b6da62a6",
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 56.4
},
{
"skill": "nemo-mbridge-resiliency",
"evaluation_date": "2026-06-02",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"has_results": true,
"average_uplift_pct": 0.5
},
{
"skill": "nemo-relay-debug-runtime-integration",
"evaluation_date": "2026-07-30",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:6c296a04671b2c4dbcf71cd6635b53ccd7fdef8405ea55c6ba6de51d66a3f0e2",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 24.0
},
{
"skill": "nemo-relay-get-started",
"evaluation_date": "2026-08-12",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:d8b84368c53829b1ab95079b2968f8e667766a8d13577bc10331d59a90e243ae",
"validation_status": null,
"tasks": 15,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 33.2
},
{
"skill": "nemo-relay-install",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 10,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 36.3
},
{
"skill": "nemo-relay-instrument-calls",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 31.3
},
{
"skill": "nemo-relay-instrument-context-isolation",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 35.7
},
{
"skill": "nemo-relay-instrument-typed-wrappers",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 32.6
},
{
"skill": "nemo-relay-migrate-from-flow",
"evaluation_date": "2026-08-12",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:738c9da141c6937b5a549c6cce2751d6b1786921b3eb39fa345e7e2ac44f8fac",
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 26.9
},
{
"skill": "nemo-relay-plugin-adaptive-tuning",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 12,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 37.7
},
{
"skill": "nemo-relay-plugin-build",
"evaluation_date": "2026-08-12",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:c8403ee1b7460312e49238c2b1fb6278c21b125e44f94b87bbe76ddf1cb710b6",
"validation_status": null,
"tasks": 7,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 32.8
},
{
"skill": "nemo-relay-plugin-observability",
"evaluation_date": "2026-07-30",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:f399331a28e80f4e118ccf4bfb6f71b126c0d88cd1103bf84c6126b50eb54d8a",
"validation_status": null,
"tasks": 20,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"has_results": true,
"average_uplift_pct": 34.2
},
{
"skill": "nemo-retriever",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"has_results": true,
"average_uplift_pct": 1.9
},
{
"skill": "nemo-rl-auto-research",
"evaluation_date": "2026-06-01",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"has_results": true,
"average_uplift_pct": 1.6
},
{
"skill": "nemo-rl-brev-etiquette",
"evaluation_date": "2026-06-01",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"has_results": true,
"average_uplift_pct": 4.5
},
{
"skill": "nemo-rl-docs",
"evaluation_date": "2026-06-01",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"has_results": true,
"average_uplift_pct": 5.1
},
{
"skill": "nemo-rl-session-memory",
"evaluation_date": "2026-06-01",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"has_results": true,
"average_uplift_pct": 4.2
},
{
"skill": "nemoclaw-user-guide",
"evaluation_date": "2026-07-17",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemoclaw-user-guide",
"component": "NemoClaw",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "nemotron-asr-finetune",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 9,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"has_results": true,
"average_uplift_pct": 28.0
},
{
"skill": "nemotron-customize",
"evaluation_date": "2026-06-04",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 8,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"has_results": true,
"average_uplift_pct": 18.2
},
{
"skill": "nemotron-policy-generator",
"evaluation_date": "2026-06-03",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 11,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"has_results": true,
"average_uplift_pct": 7.1
},
{
"skill": "nemotron-retrieval-recipes",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 40,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"has_results": true,
"average_uplift_pct": 30.2
},
{
"skill": "nemotron-speech",
"evaluation_date": "2026-08-20",
"profile": null,
"environment": "local",
"evaluator_version": "1.3.2",
"dataset_digest": "sha256:7da18a129d0ad5efdc392d764333668f68f9acf996152684bc2464427afdb20f",
"validation_status": "passed",
"tasks": 18,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"has_results": true,
"average_uplift_pct": 22.3
},
{
"skill": "nv-generate-ct-rflow",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 28.7
},
{
"skill": "nv-generate-mr",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 14.2
},
{
"skill": "nv-generate-mr-brain",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 40.5
},
{
"skill": "nv-generate-mr-brain-finetune",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 41.6
},
{
"skill": "nv-generate-vae-finetune",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 36.5
},
{
"skill": "nv-reason-cxr",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 40.4
},
{
"skill": "nv-segment-ct",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 28.4
},
{
"skill": "nv-segment-ct-finetune",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 39.8
},
{
"skill": "nv-segment-ctmr",
"evaluation_date": "2026-07-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"has_results": true,
"average_uplift_pct": 29.5
},
{
"skill": "nvidia-skill-finder",
"evaluation_date": "2026-07-24",
"profile": null,
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 17,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "nvidia-skill-finder",
"component": "NVIDIA Skills",
"has_results": true,
"average_uplift_pct": 32.8
},
{
"skill": "omniverse-cad-to-simready",
"evaluation_date": "2026-08-12",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:ce412af18bb5e516371ec07ea8752e4043a1d00edd007b505e30f659b8d92047",
"validation_status": null,
"tasks": 8,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "omniverse-cad-to-simready",
"component": "Physical AI",
"has_results": true,
"average_uplift_pct": 33.4
},
{
"skill": "omniverse-realtime-viewer",
"evaluation_date": "2026-05-28",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "omniverse-realtime-viewer",
"component": "Physical AI",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "omniverse-usd-performance-tuning",
"evaluation_date": "2026-07-30",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.0",
"dataset_digest": "sha256:137b9ca69a3516a7fffcdfff70f950f2851e5c425ccd4d9be11fb4ee4161ff8c",
"validation_status": null,
"tasks": 9,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "omniverse-usd-performance-tuning",
"component": "Physical AI",
"has_results": true,
"average_uplift_pct": 38.6
},
{
"skill": "paidf-anomalygen",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 63.5
},
{
"skill": "physical-ai-defect-image-generation",
"evaluation_date": "2026-06-23",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"has_results": true,
"average_uplift_pct": 46.3
},
{
"skill": "physical-ai-infrastructure-setup-and-resilient-scaling",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "physical-ai-infrastructure-setup-and-resilient-scaling",
"component": "Physical AI",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "physical-ai-neural-reconstruction",
"evaluation_date": "2026-07-16",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"has_results": true,
"average_uplift_pct": 36.4
},
{
"skill": "physical-ai-people-attribute-search",
"evaluation_date": "2026-07-02",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 10,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"has_results": true,
"average_uplift_pct": 1.2
},
{
"skill": "physical-ai-video-data-augmentation",
"evaluation_date": "2026-07-01",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"has_results": true,
"average_uplift_pct": 60.8
},
{
"skill": "physicsnemo-discover",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 4,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"has_results": true,
"average_uplift_pct": 6.4
},
{
"skill": "physicsnemo-shard-tensor",
"evaluation_date": "2026-08-05",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.0.0",
"dataset_digest": "sha256:dd5c60518eac8911bdb11a1e2b2eb755fb9faa4bf701c996e6334702ab7a589c",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"has_results": true,
"average_uplift_pct": 27.3
},
{
"skill": "portfolio-optimization",
"evaluation_date": "2026-07-30",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:93d7d58ac9dd6ef70bbdea737d6c97c7ff428b7568edfff7aa9028ec3966bc78",
"validation_status": null,
"tasks": 4,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"has_results": true,
"average_uplift_pct": 17.1
},
{
"skill": "rag-blueprint",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "rag-blueprint",
"component": "RAG Blueprint",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "rag-eval",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "FAIL",
"agents": [],
"catalog_dir": "rag-eval",
"component": "RAG Blueprint",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "rag-perf",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": null,
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": null,
"attempts_per_task": null,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [],
"catalog_dir": "rag-perf",
"component": "RAG Blueprint",
"has_results": false,
"average_uplift_pct": null
},
{
"skill": "rtvi-cv-customize-model",
"evaluation_date": "2026-07-31",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:7ebbe050322d781bde28f4d07a79f421dbf70294c37c25522c802460d5f12921",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "rtvi-cv-customize-model",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "rtvi-cv-scaffold-vss-service",
"evaluation_date": "2026-07-31",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:401fba496bce5a42048bf6833a613cb923e36ddf8de7626d2efcbbfbe5c0d0fb",
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 38.2
},
{
"skill": "rtvi-vlm-customize-model",
"evaluation_date": "2026-07-31",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "0.9.2",
"dataset_digest": "sha256:be4b29c85fb10f95f92e1e1cac28bc91a1a7b65cab18e3223986a0e52a66fb20",
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"has_results": true,
"average_uplift_pct": 35.6
},
{
"skill": "skill-card-generator",
"evaluation_date": "2026-06-30",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 11,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"has_results": true,
"average_uplift_pct": 21.5
},
{
"skill": "tao-analyze-changenet-rca",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 64.9
},
{
"skill": "tao-analyze-gaps-visual-changenet",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 45.0
},
{
"skill": "tao-analyze-gaps-vlm-bcq",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 54.3
},
{
"skill": "tao-convert-dataset-format",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 38.4
},
{
"skill": "tao-finetune-clip",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 64.0
},
{
"skill": "tao-finetune-cosmos-embed",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 63.6
},
{
"skill": "tao-finetune-cosmos-reason",
"evaluation_date": "2026-06-29",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 36.6
},
{
"skill": "tao-finetune-huggingface-model",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 65.2
},
{
"skill": "tao-generate-image-grounding",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 63.7
},
{
"skill": "tao-generate-referring-expressions",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 61.4
},
{
"skill": "tao-generate-video-reasoning-annotations",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 64.1
},
{
"skill": "tao-launch-workflow",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 63.8
},
{
"skill": "tao-list-capabilities",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 63.8
},
{
"skill": "tao-mine-aoi-images",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 29.4
},
{
"skill": "tao-port-huggingface-model",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 62.2
},
{
"skill": "tao-route-visual-changenet-samples",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 66.2
},
{
"skill": "tao-run-automl",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 57.4
},
{
"skill": "tao-run-automl-deft-pipeline",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 54.8
},
{
"skill": "tao-run-deft-aoi",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 55.7
},
{
"skill": "tao-run-inference-service",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 60.2
},
{
"skill": "tao-run-on-brev",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 10.4
},
{
"skill": "tao-run-on-docker",
"evaluation_date": "2026-07-21",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 69.5
},
{
"skill": "tao-run-on-kubernetes",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 36.1
},
{
"skill": "tao-run-on-local-docker",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 41.2
},
{
"skill": "tao-run-on-slurm",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 63.0
},
{
"skill": "tao-run-platform",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 53.5
},
{
"skill": "tao-setup-nvidia-gpu-host",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 43.0
},
{
"skill": "tao-train-action-recognition",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 5.8
},
{
"skill": "tao-train-bevfusion",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 24.4
},
{
"skill": "tao-train-centerpose",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 30.8
},
{
"skill": "tao-train-deformable-detr",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 19.2
},
{
"skill": "tao-train-depth-anything-v2",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 34.4
},
{
"skill": "tao-train-dino",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 54.3
},
{
"skill": "tao-train-fast-foundation-stereo",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 30.2
},
{
"skill": "tao-train-foundation-stereo",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 65.7
},
{
"skill": "tao-train-grounding-dino",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 38.6
},
{
"skill": "tao-train-image-classification",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 67.6
},
{
"skill": "tao-train-mask-auto-encoder",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 51.3
},
{
"skill": "tao-train-mask-auto-label",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 40.2
},
{
"skill": "tao-train-mask-grounding-dino",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 19.1
},
{
"skill": "tao-train-mask2former",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 58.1
},
{
"skill": "tao-train-metric-learning-recognition",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 21.4
},
{
"skill": "tao-train-nvdinov2",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 53.3
},
{
"skill": "tao-train-nvpanoptix3d",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 37.6
},
{
"skill": "tao-train-ocdnet",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 49.3
},
{
"skill": "tao-train-ocrnet",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 12.8
},
{
"skill": "tao-train-oneformer",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 64.3
},
{
"skill": "tao-train-optical-inspection",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 19.6
},
{
"skill": "tao-train-pointpillars",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 0.6
},
{
"skill": "tao-train-pose-classification",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 64.6
},
{
"skill": "tao-train-reid",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 23.1
},
{
"skill": "tao-train-rtdetr",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 55.0
},
{
"skill": "tao-train-segformer",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 22.8
},
{
"skill": "tao-train-single-step",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 31.5
},
{
"skill": "tao-train-sparse4d",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 45.8
},
{
"skill": "tao-train-visual-changenet",
"evaluation_date": "2026-06-20",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 50.0
},
{
"skill": "tao-validate-dataset-format",
"evaluation_date": "2026-06-22",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"has_results": true,
"average_uplift_pct": 46.0
},
{
"skill": "tilegym-adding-cutile-kernel",
"evaluation_date": "2026-05-29",
"profile": "external",
"environment": "local",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 1.1
},
{
"skill": "tilegym-converting-cutile-to-julia",
"evaluation_date": "2026-06-10",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "FAIL",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 11.8
},
{
"skill": "tilegym-converting-cutile-to-triton",
"evaluation_date": "2026-07-14",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 10.9
},
{
"skill": "tilegym-cutile-autotuning",
"evaluation_date": "2026-06-10",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "FAIL",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 8.5
},
{
"skill": "tilegym-cutile-python",
"evaluation_date": "2026-06-08",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 2,
"pass_threshold_pct": 50.0,
"verdict": "FAIL",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 13.6
},
{
"skill": "tilegym-improve-cutile-kernel-perf",
"evaluation_date": "2026-07-27",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [],
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 9.8
},
{
"skill": "tilegym-monkey-patch-kernels-to-transformers",
"evaluation_date": "2026-07-13",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"has_results": true,
"average_uplift_pct": 12.0
},
{
"skill": "vss-ask-video",
"evaluation_date": "2026-06-09",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 20.0
},
{
"skill": "vss-deploy-dense-captioning",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 35.5
},
{
"skill": "vss-deploy-detection-tracking-2d",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 22.5
},
{
"skill": "vss-deploy-detection-tracking-3d",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 42.1
},
{
"skill": "vss-deploy-profile",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 5,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 49.8
},
{
"skill": "vss-deploy-video-embedding",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 24.1
},
{
"skill": "vss-generate-video-calibration",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 23.2
},
{
"skill": "vss-generate-video-report",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 3,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 15.8
},
{
"skill": "vss-manage-alerts",
"evaluation_date": "2026-06-14",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 7,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 30.6
},
{
"skill": "vss-manage-video-io-storage",
"evaluation_date": "2026-06-09",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 2,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 34.4
},
{
"skill": "vss-query-analytics",
"evaluation_date": "2026-06-09",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 22.4
},
{
"skill": "vss-search-archive",
"evaluation_date": "2026-06-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 41.3
},
{
"skill": "vss-setup-behavior-analytics",
"evaluation_date": "2026-06-10",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 46.7
},
{
"skill": "vss-setup-video-analytics-api",
"evaluation_date": "2026-06-09",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 44.7
},
{
"skill": "vss-summarize-video",
"evaluation_date": "2026-07-15",
"profile": "external",
"environment": "astra-sandbox",
"evaluator_version": null,
"dataset_digest": null,
"validation_status": null,
"tasks": 1,
"attempts_per_task": 1,
"pass_threshold_pct": 50.0,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"has_results": true,
"average_uplift_pct": 37.2
},
{
"skill": "warp-compile-time-optimizer",
"evaluation_date": "2026-08-13",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:7f1d9e790399798ffbe028ff0ae8cdcf78cc28e4b304a09125ad6da274002c60",
"validation_status": null,
"tasks": 18,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "warp-compile-time-optimizer",
"component": "Warp",
"has_results": true,
"average_uplift_pct": 7.0
},
{
"skill": "warp-debug-gradients",
"evaluation_date": "2026-08-12",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.4",
"dataset_digest": "sha256:000033468255fb96783d760282543223f1de0f819410f680afa302b24fa1d4f1",
"validation_status": null,
"tasks": 6,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"has_results": true,
"average_uplift_pct": 12.5
},
{
"skill": "warp-eval",
"evaluation_date": "2026-08-11",
"profile": null,
"environment": "k8s-sandbox",
"evaluator_version": "1.2.0",
"dataset_digest": "sha256:3cef7eea3a63f09e3fd21bc5026ce689b012891fadbaed79881c8766b7e5640c",
"validation_status": null,
"tasks": 22,
"attempts_per_task": 1,
"pass_threshold_pct": null,
"verdict": "PASS",
"agents": [
"claude-code",
"codex"
],
"catalog_dir": "warp-eval",
"component": "Warp",
"has_results": true,
"average_uplift_pct": 22.8
}
],
"results": [
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 61.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "accelerated-computing-cudf",
"component": "cuDF",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "aiq-deploy",
"component": "AIQ",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 54.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 43.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 33.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 38.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 43.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 47.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "aiq-research",
"component": "AIQ",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 43.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": -8.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 51.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 53.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "amc-run-rtsp-calibration",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "amc-run-sample-calibration",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 57.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 38.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "amc-run-video-calibration",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "amc-setup-calibration-stack",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "cudaq-guide",
"component": "CUDA-Q",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 42.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "cuopt-developer",
"component": "cuOpt",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 74.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "cuopt-install",
"component": "cuOpt",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "cuopt-multi-objective-exploration",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "cuopt-numerical-optimization-api",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 94.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "cuopt-numerical-optimization-formulation",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "cuopt-routing-api-python",
"component": "cuOpt",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 56.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "cuopt-server-api-python",
"component": "cuOpt",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "cupynumeric-hdf5",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 24,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 24,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 24,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 24,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 24,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 24,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 24,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 24,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 24,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 74.0
},
{
"catalog_dir": "cupynumeric-install",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 24,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 27,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 27,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 27,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 27,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 27,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 27,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 27,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 27,
"agent": "codex",
"score_pct": 64.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 27,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "cupynumeric-migration-readiness",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 27,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 7,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Security",
"num": 7,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 7,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Correctness",
"num": 7,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 7,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Discoverability",
"num": 7,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 7,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Effectiveness",
"num": 7,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 7,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "cupynumeric-parallel-data-load",
"component": "cuPyNumeric",
"dimension": "Efficiency",
"num": 7,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": -30.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "dali-dynamic-mode",
"component": "DALI",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "data-designer",
"component": "Data Designer",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Security",
"num": 7,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Security",
"num": 7,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": -7.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Correctness",
"num": 7,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Correctness",
"num": 7,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 7,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 7,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 7,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 7,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 7,
"agent": "claude-code",
"score_pct": 65.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "deepstream-dev",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 7,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "deepstream-import-vision-model",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 53.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "deepstream-profile-pipeline",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 53.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "deepstream-run-mv3dt",
"component": "DeepStream",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "deepstream-run-mv3dt",
"component": "DeepStream",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "deepstream-run-mv3dt",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "deepstream-run-mv3dt",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "deepstream-run-mv3dt",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "deepstream-sop",
"component": "DeepStream",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "dicom-metadata-extract",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 46.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 59.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 56.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "dicom-series-preflight",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 46.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "dicom-series-to-volume",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 57.0,
"uplift_pct": -8.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 58.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "digital-health-clinical-asr-build",
"component": "Digital Health",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 53.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 49.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 44.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "digital-health-clinical-asr-eval",
"component": "Digital Health",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 57.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 56.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 47.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "digital-health-clinical-asr-finetune",
"component": "Digital Health",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "doca-aes-gcm",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "doca-argp",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 74.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-argus",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-bare-metal-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "doca-bench",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "doca-bench-extension",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "doca-bf3-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "doca-bf4-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "doca-caps",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "doca-collectx-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "doca-comch",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "doca-comm-channel-admin",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 74.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-common",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-compress",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-container-deployment",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-debug",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-devemu",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-dma",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 79.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "doca-dms",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "doca-dpa",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-dpa-hl-tracer",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-dpdk-bridge",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-erasure-coding",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-eth",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-firefly",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "doca-flow",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "doca-flow-dpa-perf",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-flow-dpa-provider",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-flow-grpc-server",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "doca-flow-perf",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "doca-flow-tune",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "doca-gpi",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-gpunetio",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-bw",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-gpunetio-ib-write-lat",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "doca-hardware-safety",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-mgmt",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-pcc",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-pcc-counters",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "doca-pcc-ztr-rttcc-algo",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-programming-guide",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-public-knowledge-map",
"component": "DOCA",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-rdma",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "doca-rdmi",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "doca-rmax",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "doca-setup",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-sha",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-sha-offload-engine",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-socket-relay",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-spcx-cc",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-sta",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-structured-tools-contract",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-telemetry",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "doca-telemetry-exporter",
"component": "DOCA",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 74.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "doca-telemetry-utils",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 58.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "doca-upgrade",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 45.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "doca-urom",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "doca-urom-svc",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-verbs",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "doca-version",
"component": "DOCA",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": -33.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "earth2studio-create-diagnostic",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 57.0,
"uplift_pct": -9.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 47.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 58.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "earth2studio-data-fetch",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 37.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 51.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "earth2studio-discover",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 36.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "earth2studio-install",
"component": "Earth2Studio",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 43.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "holohub-app-lifecycle",
"component": "HoloHub",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": -20.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": -9.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "holohub-debug-build-run",
"component": "HoloHub",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 40.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": -11.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "holohub-module-lifecycle",
"component": "HoloHub",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 33.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "holoscan-install-conda",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": -12.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "holoscan-install-container",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "holoscan-install-debian",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "holoscan-install-source",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": -19.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": -6.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "holoscan-install-wheel",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": -50.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "holoscan-setup",
"component": "Holoscan SDK",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 59.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-app",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "hsb-flash",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 63.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 59.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "hsb-ip-create-top",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 64.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "hsb-ip-def",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 55.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "hsb-ip-packetizer",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": -50.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": -100.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -68.0
},
{
"catalog_dir": "hsb-setup",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": -5.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "hsb-test",
"component": "Holoscan Sensor Bridge",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "i4h-catheter-navigation-digital-twin",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 25.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 38.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "i4h-catheter-navigation-render-drr",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 35.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 46.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-catheter-navigation-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "i4h-catheter-navigation-smoke",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "i4h-catheter-navigation-viewport",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 44.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 33.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "i4h-lerobot-viz",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 49.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-workflow",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "i4h-workflow-create",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 45.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "i4h-workflow-dataset-annotate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 31.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 52.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "i4h-workflow-dataset-convert",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 45.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 27.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "i4h-workflow-dataset-mimic",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 41.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 41.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "i4h-workflow-dataset-replay",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 63.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 46.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "i4h-workflow-dataset-teleop",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 64.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": -50.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 45.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "i4h-workflow-e2e",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 49.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "i4h-workflow-finetune",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": -17.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 36.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "i4h-workflow-scene-edit",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 63.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "i4h-workflow-setup",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": -6.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 51.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "i4h-workflow-validate",
"component": "Isaac for Healthcare Workflows",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 63.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-build-source",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 44.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 30.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 49.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-customize-camera",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "jetson-customize-clocks",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 94.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 39.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 49.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "jetson-customize-fan",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 94.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 44.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 43.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "jetson-customize-mgbe",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 51.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "jetson-customize-nvpmodel",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 39.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 54.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 59.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-customize-pcie",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "jetson-customize-pinmux",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "jetson-customize-uphy",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 52.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 63.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 53.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-customize-usb",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 51.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-derive-carrier",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 63.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "jetson-diagnostic",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 42.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 17.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 31.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 25.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "jetson-download-bsp",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "jetson-flash-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 41.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 49.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "jetson-generate-kb",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 61.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 51.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "jetson-headless-mode",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 49.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-inference-mem-tune",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-init-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 35.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "jetson-init-source",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 15.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 57.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": -26.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 47.0,
"uplift_pct": -8.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 25.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -6.0
},
{
"catalog_dir": "jetson-init-target",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 61.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 49.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 32.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 49.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "jetson-link-docs",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": -12.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 52.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "jetson-llm-benchmark",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-llm-serve",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 69.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 41.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 47.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "jetson-memory-audit",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "jetson-optimize-memory",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 41.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 54.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "jetson-package",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 45.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 58.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-print-bsp-info",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 59.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "jetson-print-device-info",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 53.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 56.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "jetson-promote-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 44.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 43.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 52.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "jetson-quick-start",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 65.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 55.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 53.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "jetson-set-target",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-speculative-decoding",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 45.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 30.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 24.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 24.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 53.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "jetson-validate-image",
"component": "Jetson BSP",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 35.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "jetson-video-benchmark",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": -17.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "jetson-video-capability",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 30.0,
"uplift_pct": -10.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 30.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "jetson-video-pipeline",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 89.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "jetson-video-recipe",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 94.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 46.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 46.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "jetson-video-setup",
"component": "Jetson Device",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "launch-nemo-rl",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "nemo-automodel-distributed-training",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 73.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-automodel-launcher-config",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "nemo-automodel-model-onboarding",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "nemo-automodel-recipe-development",
"component": "NeMo AutoModel",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": -25.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "nemo-fabric-build-adapter",
"component": "NeMo Fabric",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 40.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "nemo-fabric-integrate",
"component": "NeMo Fabric",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 95.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "nemo-mbridge-mlm-bridge-training",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-multi-node-slurm",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 87.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-mbridge-perf-activation-recompute",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -5.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cpu-offloading",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "nemo-mbridge-perf-cuda-graphs",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 92.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-expert-parallel-overlap",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-hierarchical-context-parallel",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-megatron-fsdp",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "nemo-mbridge-perf-memory-tuning",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 87.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 87.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-dispatcher-selection",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-hardware-configs",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 92.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": -12.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-long-context",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": -18.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-optimization-workflow",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 94.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-moe-vlm-training",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-parallelism-strategies",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 45.0,
"uplift_pct": -29.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "nemo-mbridge-perf-sequence-packing",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-perf-tp-dp-comm-overlap",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 55.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 98.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-mbridge-recipe-recommender",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-mbridge-resiliency",
"component": "NeMo MBridge",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 56.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "nemo-relay-debug-runtime-integration",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -7.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 65.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nemo-relay-get-started",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "nemo-relay-install",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "nemo-relay-instrument-calls",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 65.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "nemo-relay-instrument-context-isolation",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-relay-instrument-typed-wrappers",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 28.0,
"uplift_pct": -22.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 29.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "nemo-relay-migrate-from-flow",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "nemo-relay-plugin-adaptive-tuning",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "nemo-relay-plugin-build",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": -5.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemo-relay-plugin-observability",
"component": "NeMo Relay",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 45.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 47.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-retriever",
"component": "NeMo Retriever",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": -9.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": -9.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": -6.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 57.0,
"uplift_pct": -5.0
},
{
"catalog_dir": "nemo-rl-auto-research",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-rl-brev-etiquette",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 81.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "nemo-rl-docs",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": -5.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "nemo-rl-session-memory",
"component": "NeMo-RL",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "nemotron-asr-finetune",
"component": "Nemotron Speech",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 63.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nemotron-customize",
"component": "Nemotron",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 54.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 64.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "nemotron-policy-generator",
"component": "Nemotron",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "nemotron-retrieval-recipes",
"component": "Nemotron",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": -6.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "nemotron-speech",
"component": "Nemotron Speech",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 64.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 38.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 36.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nv-generate-ct-rflow",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 18.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 57.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "nv-generate-mr",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 74.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 13.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 40.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "nv-generate-mr-brain",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 58.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "nv-generate-mr-brain-finetune",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 57.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 69.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nv-generate-vae-finetune",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 54.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "nv-reason-cxr",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 21.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "nv-segment-ct",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 69.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "nv-segment-ct-finetune",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 63.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 39.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 54.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 42.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "nv-segment-ctmr",
"component": "Medical AI Skills",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "nvidia-skill-finder",
"component": "NVIDIA Skills",
"dimension": "Security",
"num": 17,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "nvidia-skill-finder",
"component": "NVIDIA Skills",
"dimension": "Correctness",
"num": 17,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "nvidia-skill-finder",
"component": "NVIDIA Skills",
"dimension": "Discoverability",
"num": 17,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "nvidia-skill-finder",
"component": "NVIDIA Skills",
"dimension": "Effectiveness",
"num": 17,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "nvidia-skill-finder",
"component": "NVIDIA Skills",
"dimension": "Efficiency",
"num": 17,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "omniverse-cad-to-simready",
"component": "Physical AI",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "omniverse-cad-to-simready",
"component": "Physical AI",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "omniverse-cad-to-simready",
"component": "Physical AI",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "omniverse-cad-to-simready",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "omniverse-cad-to-simready",
"component": "Physical AI",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "omniverse-usd-performance-tuning",
"component": "Physical AI",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "omniverse-usd-performance-tuning",
"component": "Physical AI",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "omniverse-usd-performance-tuning",
"component": "Physical AI",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "omniverse-usd-performance-tuning",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 63.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "omniverse-usd-performance-tuning",
"component": "Physical AI",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "paidf-anomalygen",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 71.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 79.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 77.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "physical-ai-defect-image-generation",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 69.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Security",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Security",
"num": 4,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Correctness",
"num": 4,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Correctness",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 4,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 4,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 4,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 4,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 4,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "physical-ai-neural-reconstruction",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 4,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": -1.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 82.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-people-attribute-search",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "physical-ai-video-data-augmentation",
"component": "Physical AI",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": -9.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": -5.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "physicsnemo-discover",
"component": "PhysicsNeMo",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 73.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "physicsnemo-shard-tensor",
"component": "PhysicsNeMo",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 23.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 52.0,
"uplift_pct": 2.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 41.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "portfolio-optimization",
"component": "Portfolio Optimization",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 29.0
},
{
"catalog_dir": "rtvi-cv-customize-model",
"component": "DeepStream",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": -33.0
},
{
"catalog_dir": "rtvi-cv-customize-model",
"component": "DeepStream",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "rtvi-cv-customize-model",
"component": "DeepStream",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "rtvi-cv-customize-model",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "rtvi-cv-customize-model",
"component": "DeepStream",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 79.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "rtvi-cv-scaffold-vss-service",
"component": "DeepStream",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 75.0,
"uplift_pct": -8.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 63.0
},
{
"catalog_dir": "rtvi-vlm-customize-model",
"component": "DeepStream",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 91.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 55.0,
"uplift_pct": -45.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 37.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 69.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "skill-card-generator",
"component": "Skill Card Generator",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 89.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-analyze-changenet-rca",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-analyze-gaps-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 34.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 24.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "tao-analyze-gaps-vlm-bcq",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 10.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 42.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-convert-dataset-format",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "tao-finetune-clip",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "tao-finetune-cosmos-embed",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "tao-finetune-cosmos-reason",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-finetune-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "tao-generate-image-grounding",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 82.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 63.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "tao-generate-referring-expressions",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "tao-generate-video-reasoning-annotations",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-launch-workflow",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 56.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-list-capabilities",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-mine-aoi-images",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-port-huggingface-model",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 92.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "tao-route-visual-changenet-samples",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 72.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-automl",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "tao-run-automl-deft-pipeline",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 42.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "tao-run-deft-aoi",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 40.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-inference-service",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 48.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-run-on-brev",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 94.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "tao-run-on-docker",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "tao-run-on-kubernetes",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 47.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 60.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "tao-run-on-local-docker",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "tao-run-on-slurm",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-run-platform",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 59.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-setup-nvidia-gpu-host",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-action-recognition",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": -10.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-bevfusion",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "tao-train-centerpose",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "tao-train-deformable-detr",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 48.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "tao-train-depth-anything-v2",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 72.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "tao-train-dino",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 60.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 38.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-fast-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "tao-train-foundation-stereo",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 76.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "tao-train-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 92.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "tao-train-image-classification",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-mask-auto-encoder",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 66.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-auto-label",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": -4.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "tao-train-mask-grounding-dino",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 77.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 70.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "tao-train-mask2former",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "tao-train-metric-learning-recognition",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 85.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 58.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "tao-train-nvdinov2",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 61.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 10.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 42.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 76.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-nvpanoptix3d",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 77.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 34.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 24.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "tao-train-ocdnet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-ocrnet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 83.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 86.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 67.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "tao-train-oneformer",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 44.0,
"uplift_pct": -34.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-optical-inspection",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pointpillars",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 88.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 75.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "tao-train-pose-classification",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 34.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 44.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 24.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "tao-train-reid",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "tao-train-rtdetr",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 20.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 85.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 82.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 32.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 35.0
},
{
"catalog_dir": "tao-train-segformer",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 84.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 34.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 24.0,
"uplift_pct": -3.0
},
{
"catalog_dir": "tao-train-single-step",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 92.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 90.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 74.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 79.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "tao-train-sparse4d",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 77.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-train-visual-changenet",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 97.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 80.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tao-validate-dataset-format",
"component": "TAO Toolkit",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Security",
"num": 8,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Security",
"num": 8,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Correctness",
"num": 8,
"agent": "claude-code",
"score_pct": 93.0,
"uplift_pct": -2.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Correctness",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 3.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Discoverability",
"num": 8,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Discoverability",
"num": 8,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 8,
"agent": "claude-code",
"score_pct": 95.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 8,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Efficiency",
"num": 8,
"agent": "claude-code",
"score_pct": 77.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "tilegym-adding-cutile-kernel",
"component": "TileGym",
"dimension": "Efficiency",
"num": 8,
"agent": "codex",
"score_pct": 85.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-julia",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 98.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "tilegym-converting-cutile-to-triton",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tilegym-cutile-autotuning",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Security",
"num": 6,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Security",
"num": 6,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Correctness",
"num": 6,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Correctness",
"num": 6,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Discoverability",
"num": 6,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Discoverability",
"num": 6,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 6,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 1.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 6,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Efficiency",
"num": 6,
"agent": "claude-code",
"score_pct": 78.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "tilegym-cutile-python",
"component": "TileGym",
"dimension": "Efficiency",
"num": 6,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 16.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 18.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 87.0,
"uplift_pct": 4.0
},
{
"catalog_dir": "tilegym-improve-cutile-kernel-perf",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 14.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 99.0,
"uplift_pct": 17.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "tilegym-monkey-patch-kernels-to-transformers",
"component": "TileGym",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 7.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-ask-video",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 25.0,
"uplift_pct": -75.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 53.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 49.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 55.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 57.0,
"uplift_pct": 54.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 59.0
},
{
"catalog_dir": "vss-deploy-dense-captioning",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 97.0,
"uplift_pct": 41.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 52.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-2d",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": 52.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 87.0,
"uplift_pct": 47.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 74.0,
"uplift_pct": 68.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 63.0,
"uplift_pct": 48.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "vss-deploy-detection-tracking-3d",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 5,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 5,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 10.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 5,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 81.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 5,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 5,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 70.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 5,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 5,
"agent": "claude-code",
"score_pct": 80.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 5,
"agent": "codex",
"score_pct": 70.0,
"uplift_pct": 64.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 5,
"agent": "claude-code",
"score_pct": 71.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "vss-deploy-profile",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 5,
"agent": "codex",
"score_pct": 88.0,
"uplift_pct": 40.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 66.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "vss-deploy-video-embedding",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 68.0,
"uplift_pct": 42.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 73.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 31.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 33.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 64.0,
"uplift_pct": 12.0
},
{
"catalog_dir": "vss-generate-video-calibration",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 3,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 3,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 3,
"agent": "claude-code",
"score_pct": 52.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 3,
"agent": "codex",
"score_pct": 23.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 3,
"agent": "claude-code",
"score_pct": 8.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 3,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 3,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 3,
"agent": "codex",
"score_pct": 34.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 3,
"agent": "claude-code",
"score_pct": 26.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "vss-generate-video-report",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 3,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 7,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 7,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": -7.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 7,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 7,
"agent": "codex",
"score_pct": 78.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 7,
"agent": "claude-code",
"score_pct": 99.0,
"uplift_pct": 55.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 7,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 7,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 7,
"agent": "codex",
"score_pct": 51.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 7,
"agent": "claude-code",
"score_pct": 89.0,
"uplift_pct": 51.0
},
{
"catalog_dir": "vss-manage-alerts",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 7,
"agent": "codex",
"score_pct": 80.0,
"uplift_pct": 22.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 2,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "claude-code",
"score_pct": 90.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 2,
"agent": "codex",
"score_pct": 92.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 69.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 2,
"agent": "codex",
"score_pct": 71.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "claude-code",
"score_pct": 60.0,
"uplift_pct": 38.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 2,
"agent": "codex",
"score_pct": 66.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "claude-code",
"score_pct": 83.0,
"uplift_pct": 61.0
},
{
"catalog_dir": "vss-manage-video-io-storage",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 2,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 19.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 62.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 62.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 27.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-query-analytics",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": -0.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 97.0,
"uplift_pct": 43.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 68.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 62.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 72.0
},
{
"catalog_dir": "vss-search-archive",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 81.0,
"uplift_pct": 39.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 0.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 100.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 50.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "vss-setup-behavior-analytics",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 28.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 27.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 88.0,
"uplift_pct": 78.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 44.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 67.0
},
{
"catalog_dir": "vss-setup-video-analytics-api",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Security",
"num": 1,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 75.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Correctness",
"num": 1,
"agent": "codex",
"score_pct": 67.0,
"uplift_pct": 45.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 50.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Discoverability",
"num": 1,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 46.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "claude-code",
"score_pct": 76.0,
"uplift_pct": 56.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Effectiveness",
"num": 1,
"agent": "codex",
"score_pct": 58.0,
"uplift_pct": 34.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "claude-code",
"score_pct": 92.0,
"uplift_pct": 36.0
},
{
"catalog_dir": "vss-summarize-video",
"component": "Video Search and Summarization",
"dimension": "Efficiency",
"num": 1,
"agent": "codex",
"score_pct": 77.0,
"uplift_pct": 30.0
},
{
"catalog_dir": "warp-compile-time-optimizer",
"component": "Warp",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 11.0
},
{
"catalog_dir": "warp-compile-time-optimizer",
"component": "Warp",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 6.0
},
{
"catalog_dir": "warp-compile-time-optimizer",
"component": "Warp",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 89.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "warp-compile-time-optimizer",
"component": "Warp",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 95.0,
"uplift_pct": 5.0
},
{
"catalog_dir": "warp-compile-time-optimizer",
"component": "Warp",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 96.0,
"uplift_pct": 8.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 83.0,
"uplift_pct": -17.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 84.0,
"uplift_pct": 26.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 84.0,
"uplift_pct": 24.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 94.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 90.0,
"uplift_pct": 13.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "warp-debug-gradients",
"component": "Warp",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 93.0,
"uplift_pct": 33.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Security",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 0.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Security",
"num": null,
"agent": "codex",
"score_pct": 100.0,
"uplift_pct": 9.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Correctness",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 15.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Correctness",
"num": null,
"agent": "codex",
"score_pct": 94.0,
"uplift_pct": 20.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Discoverability",
"num": null,
"agent": "claude-code",
"score_pct": 100.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Discoverability",
"num": null,
"agent": "codex",
"score_pct": 91.0,
"uplift_pct": 31.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Effectiveness",
"num": null,
"agent": "claude-code",
"score_pct": 86.0,
"uplift_pct": 25.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Effectiveness",
"num": null,
"agent": "codex",
"score_pct": 86.0,
"uplift_pct": 28.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Efficiency",
"num": null,
"agent": "claude-code",
"score_pct": 96.0,
"uplift_pct": 32.0
},
{
"catalog_dir": "warp-eval",
"component": "Warp",
"dimension": "Efficiency",
"num": null,
"agent": "codex",
"score_pct": 98.0,
"uplift_pct": 36.0
}
]
}