What a second review catches when it asks different questions.

The setup

A variant definition tells the analysis what one variant means: its gene, which allele carries the effect, how strong the evidence is and the wording a report uses. Candidate definitions come from public evidence (ClinVar, the GWAS Catalog, PharmGKB). An AI drafter writes each one, then a separate AI reviewer, in its own context, is told to refute the draft. Drafts that survive wait in a queue. Nothing in that queue reaches a report until a reviewer countersigns it.

By late August 1,284 drafts had passed the adversarial review and were waiting. Before approving any of them we took a sample of 150 and audited it again, with fresh AI reviewers attacking on axes the first review had not used: direction of effect, which allele is major or minor, honesty about penetrance, evidence strength against wording, and condition names copied from a submitter’s free text.

The result

Only 37.3% of drafts that had already passed one adversarial review were clean. We then audited the whole queue the same way: 113 of the 1,284 were defective, 168 needed a geneticist’s judgement and 1,003 were clean or cosmetic. The defective rate held at full scale: 8.7% in the sample, 8.8% across the queue. The share needing a geneticist did not: 23.3% in the sample, 13.1% across the queue.

What the defects were

The largest class was direction of effect: an allele that protects, scored as one that harms. One draft curated a risk allele whose GWAS effect is negative. Two others described, in their own text, an allele that reduces harm while marking it as not protective. If a definition like that went live, every carrier’s result would be inverted.

Other defects were quieter but just as real:

• a palindromic C/G site with the wrong reference base;
• a germline allele at JAK2 codon 617 conflated with the somatic V617F change at the same position;
• a benign untranslated-region variant given a hereditary angioedema label taken from a submitter’s condition text;
• an X-linked dominant condition worded as though heterozygous women were unaffected.

A mechanical sweep of all 1,284 drafts for the direction-of-effect class found 2, against roughly 40 implied by the sample. Whether an allele protects or harms is not decidable from the text alone. It takes genetics knowledge.

Why the first review missed them

Both reviews ran in independent contexts, so this was not a model marking its own homework. The first review missed these rows because of what it was asked: the defects sat on axes its checklist never covered. We did not re-run the first checklist as a control, so we cannot say what it would have found a second time. What found these defects was a different checklist.

What we changed

• Nothing was bulk-approved. Every defective draft is held with its reason recorded.
• Drafts that need genetics judgement go to a geneticist with the specific question attached.
• Our review policy now requires any second review, human or AI, to attack on different axes from the first, with both verdicts recorded.
• Letting an AI hold the second seat also requires machine-checked rules before anything goes live and a re-review of a random slice of every approved batch. Until that is built and calibrated, a person countersigns every definition.

The transferable part

If you review AI output with AI, independence is necessary but not sufficient. The questions have to differ too. What found these defects was a different checklist, not another sample of the same one.

These are reference-data drafts, not patient results: no patient data was involved, and no defective draft was approved.