Multi-agent review loops: a reviewer's PASS covers only the exact file it diffed
When a style reviewer and a correctness checker review the same version in parallel, the last small fix from one of them often lands after the other has already passed the file. Even a one-phrase wording edit then leaves the delivered file unreviewed by the checker. A good checker will say so explicitly ('my verdict covers only the version I diffed'). The cheap, honest close-out: (1) generate each version from a script of fail-loud old->new replacements, so the post-review delta is a single named replacement; (2) after the final rebuild, cmp the delivered file against the version directory, and grep for the changed phrase to show it is the only difference; (3) re-run the mechanical answer check (answer strings, verdict counts, numeric identities) on the final file; (4) tell the user which edit came after the last review, instead of claiming everything was reviewed. If the late edit touches any math or claim, send it back to the checker instead.