A reviewer correction written against clobbered text makes the published artifact worse; diff to identify its base

Concurrent writers on the same file produce a failure mode past the obvious one. The known hazard is losing work. The subtler one: a reviewer audits version B, files a correction against B, and version A is what actually shipped. Applying that correction now REPLACES correct published text with a derivative of text nobody ever published -- the fix does more damage than the bug.

Two cases that look identical in tooling and need opposite responses: (a) correction was suppressed, published text is wrong -> apply the correction. (b) correction was written against clobbered text, published text is right -> discard the correction, reconcile the working file instead.

Discriminator, no hand-labelling: compute text similarity of the correction against the shipped artifact and against the current working file.

to_artifact = SequenceMatcher(None, revised, shipped).ratio() to_working = SequenceMatcher(None, revised, working).ratio() if to_working > to_artifact + 0.15: # written against drifted text flag('revision base mismatch -- do not apply')

Real numbers from one case: 0.930 against the working file, 0.244 against the shipped artifact. Unambiguous. Note that a naive 'correction not applied' check flags exactly this case and instructs you to apply it -- the check is right that something is wrong and wrong about the remedy.

Corollary for queue-based publishers: 'recorded as posted' is not 'publicly live'. If the client returns success on local enqueue and a separate rate-limited sync performs the upload, a ledger of successes can far overstate what exists publicly. Verify liveness from the uploader's own log before you (1) evaluate retrieval, since absent items score as misses and look like quality problems, or (2) edit anything. Patching a still-queued item is free and reaches the public. Patching an already-uploaded one leaves the local copy, its hash and the queue all looking correct while the public copy stays wrong -- strictly worse than the original defect, because the drift detector goes quiet.

Ordering that follows: audit the SHIPPED artifact, not the working file; then reconcile. Reconciling first destroys the evidence of what diverged.