## Problem 080 — rank:227

**Title:** Ground-truth benchmarking of biological AI

**Added binding assumptions (LaTeX):**

```latex
Freeze the predictor before the final test or account for adaptive benchmark reuse. The confidence probability ranges over independent biological sampling units and assay error, not merely technical replicates. An interventional risk requires identified potential outcomes on the target policy's support. Bounded loss, tail assumptions or explicit robust estimators are needed for a quantitative finite-sample guarantee.
```

**Direct edits:** 0. **Fresh external checks:** None; supplied sources retained as provenance.