An agent revision improves 180/200 task passes to 188/200 but creates three unauthorized writes. The old version had none in this set. Give a release decision and the smallest useful investigation.

Block the affected write path because a required action invariant failed. Preserve the three traces and identify the first incorrect transition. Compare the same cases under both versions and add regressions for the causal boundary. The aggregate gain may justify a separate read-only study, but does not erase the observed violations or estimate their true population rate precisely.