Calibrated triage and audit of auto-admitted records

[NingyuSUN]

Part of the AI validation roadmap (docs/AI_VALIDATION_ROADMAP.md).

Goal. Make "AI does most of the work, experts review only high-risk samples" measurable: route records into auto-admit / expert review (risk-ranked) / reject, and bound the error rate of what is auto-admitted.

Done when

• Routing uses only risk signals shown to predict errors (e.g. dissenting evidence), each with its evidence
• bioevidence review audit-sample: a seeded random sample of auto-admitted records for expert audit
• Reported upper bound on the auto-admitted error rate (e.g. 0 errors in 300 audited → below 1% at 95%)
• Per-route statistics: share of records, audited error rate, share of expert time

[NingyuSUN]

Done in #43.

• Routing evidence. evaluation/llm_benchmark/risk_signals.py tests each signal that sends a record to a person. BEV004 is the only risk signal on by default. Answers it flagged were wrong far more often than the rest: 82% vs 33% held-out, 70% vs 38% on external data, 91% vs 16% in the literature pilot. The opt-in signals (BEV022, BEV025, BEV026) showed no such effect, so they stay off by default.
• bioevidence review audit-sample. Draws a seeded random sample of auto-admitted records and can mix in blind controls from the other routes. --target 0.01 audits 299 records: if none is wrong, the error rate is below 1% at 95% confidence.
• bioevidence review audit-score. Reports for each route its share of records, its audited error rate with Wilson and exact one-sided (Clopper–Pearson) bounds, and its share of expert time from minutes_spent.
• Dry run on the single-cell held-out results. No expert has audited these records yet, so the dataset authors' labels stood in for the auditor. They found 14 errors in 59 sampled records, putting the error rate below 34.6% at 95% confidence. The census over all 255 auto-admitted records is 19.6%, inside the interval.

The expert audit itself is still open and depends on #23.