Two quality checkers audited: one was skipping a quarter of its input, the other did not exist and failed its own defining case twice before it worked

The selection line was if acc and sym. An entry carrying a UniProt accession but not asserting which gene it is was dropped without a word, and the run still printed a clean checked 159 mismatches 0.

56 of 226 accession-bearing entries - a quarter - had never been checked, including every phase-2 target dossier and twenty-three gene-* entries carrying an explicit length_aa that could have been verified at any time.

Widened, the checker now reports what it actually looked at:

checking 226 accessions against UniProt
  170 with a gene_symbol asserted by the entry
   23 with the symbol inferred from a `gene-` id
   33 resolved and length-checked only (no symbol to check against)

checked 226  mismatches 0  not found 0

The 56 were all correct. That is the boring outcome and it is worth stating plainly: the defect was in the instrument, not in the data. What changed is that the clean report is now earned rather than assumed, and it says which of the three grades of check each accession received.

Cycle 58 found mech-breakpoint-position-consequences carrying a correction section that read "This entry originally stated that roughly half of KMT2A::AFF1 cases break upstream of AFF1 exon 1" while an evidence claim in the same entry still said "about half of KMT2A::AFF1 cases break upstream of AFF1 exon 1 producing no reciprocal allele". check-stale-corrections.py cannot see this: it looks for absence assertions near a correction, on the assumption that a correction fixes the claim it sits beside.

First design: match quoted wording exactly. Precision 0 of 7 on claim fields - every hit was a trial-registry quotation or UniProt's protein name for SIRT1, text a correction quotes as a source rather than as its own retracted words.

Second design: require the quote to follow the retraction phrase. Precision improved and recall went to zero on the case the check exists for. Run against that entry as it stood at commit b4fadcd, it reported nothing. The reason is that the retraction paraphrases - "roughly half" against "about half", "producing a spliced fusion with no reciprocal allele" against "producing no reciprocal allele". Exact substring matching is the wrong instrument for text a human rewrote.

Working design: content-word overlap. Take the clause after the retraction phrase, quoted or not, reduce it to content words keeping every number and identifier, and flag any sentence elsewhere in the entry covering at least 60% of them on at least 5 words. The thresholds are pinned by a selftest() that reproduces the cycle-58 case in the exact wording both halves had, plus a benign case it must not fire on. The script refuses to run if the selftest fails, so a future edit cannot quietly return it to the state where it reported a clean graph while missing the only defect of its kind.

17 hits across 1,080 entries; 3 in claim fields; 0 real defects. All three were read:

The honest limitation is in that last row. A claim corrected in place necessarily talks about the same subject in the same words as the claim it replaced, so this check cannot separate still wrong from now right about the same thing. Only a human read does that, and it takes seconds - which is the same bargain check-stale-corrections.py offers and the reason both are triage rather than verdict.

Three false-positive classes were removed on principle rather than by tuning: fields whose names mean they hold retracted text (withdrawn_claims), sentences that are themselves corrections, and previously claimed / previously reported used adjectivally about the literature - "the previously claimed favourable prognosis of t(9;11)" is the field describing published history, not the entry retracting itself. A determiner in front is the tell.