Files
theclaymethod__unslop/evals/core-thresholds.json
2026-08-04 19:05:27 -07:00

49 lines
2.0 KiB
JSON

{
"schema": "unslop-core-thresholds-v1",
"frozen_at": "2026-08-04T18:27:35-07:00",
"frozen_before": "public-holdout-v9",
"post_run_integrity_fix": {
"at": "2026-08-04T19:00:07-07:00",
"scope": "core_metrics manifest normalization and structured exact-span override only",
"reason": "Runner accepted constraint.text but scorer prompt reconstruction required constraint.description. A first post-run override inferred exact-span enforcement from policy prose; independent review rejected that heuristic, so deterministic override now requires enforcement=exact_span. These changes make v17 reproducible but not a pristine pre-frozen trial; its no-ship result is unchanged.",
"thresholds_changed": false,
"model_outputs_inspected": true
},
"frozen_inputs": {
"public_holdout_v9_sha256": "aa24a1236b8a19d2444de1b8665bceb55155a638fa073d9a155aecadaba8b0e9",
"shipping_contract_sha256": "eb1b17dd714c496fe878df88c44cf5a71cd6abbf61b01b3ab11093d4e579a959",
"core_runner_sha256": "e313f3fb3134c7179c4936f48328fddb0e16ea53044e3c9431018a5d0ad6b830",
"core_metrics_sha256": "a5676d241b44b5c20aade831b93a4ddc8b73bdae85c878ef6011ff724e7b47fa",
"scanner_sha256": "fecca26e3e41ac07c032578e9419ebd284dc91d09654a401135e612d2b00771f"
},
"with_skill": {
"minimum_detection_precision": 0.9,
"minimum_detection_recall": 0.9,
"minimum_repair_success": 0.9,
"minimum_preservation": 1.0,
"maximum_damage_rate": 0.0,
"minimum_net_improvement": 0.8,
"minimum_clean_noop_rate": 1.0
},
"comparison": {
"minimum_paired_win_rate": 0.5,
"maximum_without_skill_wins": 0,
"minimum_positive_quality_lifts": 1,
"quality_lifts": [
"detection_precision",
"detection_recall",
"repair_success",
"preservation",
"damage_rate_inverse",
"net_improvement",
"clean_noop_rate"
]
},
"required_controls": {
"shipping_validation_clean": true,
"isolated_model_workspaces": true,
"independent_blinded_judge": "gpt-5.6-sol",
"generation_model_both_arms": "gpt-5.6-luna"
}
}