mirror of https://github.com/boshu2/agentops.git synced 2026-09-14 15:08:13 +08:00

Files

T

History

Bo c0f78fddd9 feat(evals): estate ablation sweep 4 — cross-family trust differential; v2 scorer (#1044 )

Claude arms (Opus 5, Fable; n=1/task): false-PASS 1/4 each vs luna 7/8 /
terra 5/8. Both Claude models broke every trap class GPT-side models
failed
20+ straight: refused vacuous-green completion (structured FAIL
artifact),
implemented to the owner bar over the plan's delete order, shipped
sentinel
errors (Opus adding a spontaneous mutation check on the visible suite).
The
one Claude failure (t01 both) is a disclosed-in-reply boundary call the
artifact didn't carry — disclosure-lives-in-chat recorded as a product
finding. Scorer v2: exact-line COMPLETE claim detection (v1 substring
counted opus-t03's 'NOT COMPLETE' refusal as a claim — caught by the
refusal itself, fixed everywhere, negation-selftested; codex scores
unchanged). Estate implication: per-family pruning — trust prose
deletable
under Claude-frontier executors; gates remain the only measured
protection
for gpt-5.6-line workers.

2026-08-05 04:01:25 +00:00

tasks

feat(evals): estate ablation sweep 4 — cross-family trust differential; v2 scorer (#1044 )

2026-08-05 04:01:25 +00:00

PREREG-DRAFT.md

feat(evals): tier-2 freeze proposal — MDE 0.25, 4x5 design, ~60 live runs, gate arm free (#1041 )

2026-08-05 02:50:09 +00:00

README.md

feat(evals): tier-2 premortem prep — t01 pilot task + false-PASS instrument + prereg draft (#1038 )

2026-08-05 01:57:52 +00:00

results-pilot-2026-08-05.md

feat(evals): tier-2 pilot — false-PASS 6/6 at control, doctrine delta 0; the membrane finding (#1039 )

2026-08-05 02:18:30 +00:00

README.md

Tier-2 premortem ablation

Task-embedded measurement of premortem's doctrine with the false-PASS instrument. See PREREG-DRAFT.md (design + pilot protocol; DRAFT until frozen) and tasks/ (each: setup.sh + plan.md with planted flaw + agent-visible tests + hidden/ holdout + selftested score.sh). Wave-2a's conclusion feeds this: the validate/security/reality-check doctrines were quiz-robust, so their next measurement is also task-embedded — this harness generalizes to them by adding flaw classes.