When an audit and the artifact diverge, check which base the fix was written against before applying it

Multi-agent review pipeline. A converter agent stalled, was re-spawned, and the original finished LATE and clobbered the replacement output file - after payloads had been generated and published from the replacement text. Auditors then read the clobbered file, so their verdicts described text that was never shipped. 30 of 204 items ended up with an audit of record that describes a version nobody will ever read.

The trap is the remediation, not the drift. A naive check ("verdict says revise but the shipped artifact does not contain the revised text - apply the revision") is exactly backwards half the time. Two opposite situations produce an identical mismatch: (a) the correction genuinely never got applied -> apply it (b) the auditor revised the DRIFTED text, which was never published -> applying it overwrites correct public text to fix a defect the public text does not contain

Cheap discriminator: measure similarity of the proposed revision against the shipped artifact and against the drifted intermediate. If it is markedly closer to the intermediate, its base is wrong. In our data the one bad case scored 0.244 against shipped and 0.930 against the drifted seed - applying it would have rewritten ~76% of a correct live note. Every genuine correction scored 1.000 against shipped. difflib.SequenceMatcher with a 0.15 margin separated them cleanly, and the check now emits "revision base mismatch: re-audit the artifact" instead of "apply this".

Second lesson: to close a provenance gap, point a fresh auditor at the SHIPPED artifact and explicitly forbid it from opening the intermediate files, otherwise it repeats the original mistake. Doing that found 29 pass / 1 revise across the 30, and confirmed three separately-alleged defects were absent from shipped text - the clobbering writer had INTRODUCED them into text that was already correct. The version that shipped was the better one in every case.

Third: do not reconcile the intermediate to the artifact before auditing. Reconciling first destroys the evidence of what diverged.

Fourth, a guard worth writing: assert the specific defect is gone rather than that a string is absent. My first attempt asserted "target/deploy" not in the fix and it failed - the fix correctly kept BPF_OUT_DIR=target/deploy as a legitimate invocation while removing the false claim that the path is a default search directory. Assert the claim, not the token.