Add audit sampling and routing evidence (#24)

[NingyuSUN]

Implements #24: audit sampling of auto-admitted records, and evidence for which routing signals predict errors.

What it adds

bioevidence review audit-sample: a seeded random sample of one method's auto-admitted records for one use, written as a blank annotation sheet.

• --target 0.01 sizes the sample so that, if no audited record is wrong, the auto-admitted error rate is below 1% at 95% confidence. That is 299 records; a 2% target needs 149, 5% needs 59 and 10% needs 29.
• --controls mixes in records from the other routes in shuffled order, so auditors cannot tell which route a record took.
• The manifest records the seed, the size of each route and the sampled hashes. It stays with the maintainer.

bioevidence review audit-score: reports, for each route:

• its share of records;
• the audited error rate, with a Wilson interval and an exact one-sided (Clopper–Pearson) upper bound;
• the most errors the route can hold;
• its share of expert time, estimated from minutes_spent.

It refuses labels for records that were not sampled, or that were made on a different record or profile.

Routing evidence (evaluation/llm_benchmark/risk_signals.py): for each signal that sends a record to a person, how often first answers were wrong when it fired and when it did not.

Signal	Default	Wrong when flagged vs otherwise	Verdict
BEV004 contradicting evidence	on	held-out 82% vs 33%; external 70% vs 38%; literature pilot 91% vs 16%	shown
BEV022 semantic cue	opt-in	4% vs 9%	not shown
BEV025 ASCT+B conflict	opt-in	57% vs 40%	not shown
BEV026 definition contradicted	opt-in	47% vs 41%	not shown

BEV004 is the only risk signal on by default, and it holds on every split. I set the criterion after the runs: the risk ratio's 95% interval must lie above 1 on a held-out or external split.

Audit dry run (evaluation/llm_benchmark/audit_celltype.py), on the single-cell held-out results. No expert has audited these records. The dry run uses the dataset authors' labels as a stand-in auditor:

• The sample was 59 auto-admitted records and 10 controls. The packet shows each record without its route, its model or the authors' label.
• The stand-in found 14 errors in 59: 23.7%, Wilson 14.7–36.0%, upper bound 34.6%.
• The true rate over all 255 auto-admitted records is 19.6% (50/255). It falls inside the interval.
• So the route misses a 5% target. An audit is how a deployment would learn that before relying on it.

Done-when in #24

• Routing uses only risk signals shown to predict errors, each with its evidence.
• bioevidence review audit-sample draws a seeded random sample of auto-admitted records.
• An upper bound on the auto-admitted error rate is reported (0 errors in 299 audited records gives a bound below 1% at 95% confidence).
• Per-route statistics: share of records, audited error rate, and share of expert time. Expert time stays empty in the dry run because the stand-in records none.

The expert audit itself still needs experts (#23).

Checks

• ruff, mypy and pytest pass, with 98.5% coverage. tests/test_audit.py checks the bounds against the closed form and against the binomial CDF.
• mkdocs build --strict passes.
• tools/reproduce.py now has 20 steps, and all of them reproduce identically, including the 2 new ones.