PASS
atomic-calibrated-verifier
- Atomic checklist
- 100%
- Item accuracy
- 100%
- False-pass rate
- 0%
- Item Brier
- 0.001
- Paired relations
- 100%
- Max reward inflation
- -0.015
FinMirror ยท deterministic meta-evaluation
Did the judge reward the right requirement?
A zero-model-call audit that separates checklist decomposition defects from permissive verifier judgments, then tests reward behavior under atomic omission, irrelevant context, and requirement reordering.
BLOCKED
atomic-permissive-verifier
- Atomic checklist
- 100%
- Item accuracy
- 94%
- False-pass rate
- 100%
- Item Brier
- 0.059
- Paired relations
- 67%
- Max reward inflation
- +0.210
BLOCKED
collapsed-permissive-verifier
- Atomic checklist
- 0%
- Item accuracy
- 75%
- False-pass rate
- 100%
- Item Brier
- 0.232
- Paired relations
- 0%
- Max reward inflation
- +0.210
Positive and negative controls: all three systems see the same oracle requirement states. Only the atomic calibrated verifier earns the release gate; permissive scores and collapsed checklists fail for different, inspectable reasons.
Reproduce with finmirror judge-demo. Supply external learned-judge outputs with finmirror judge-audit --input audit.json.