# CAIN-A75-056 Benchmark Evidence Review

**Verdict:** VERIFIED

> Recomputes a benchmark score from its record: hashed tasks and inputs, the model and decoding settings, an independent grader with its rubric, every result, and the published failures and near misses. It does not re-run the model.

| Benchmark | Model | Verdict | Score | Passed | Failures | Near misses | Failed checks |
|---|---|---|---|---|---|---|---|
| 0 bench:multimodal-knowledge-work-sample | model:acme-vision-7 | SOUND | 0.73 | 3/5 | T2, T3 | T2 |  |

## What this proves and what it does not
- Proves: every task commits its inputs and prompt by SHA-256, comes from a test or held-out split and is not in the declared training hashes; the model, version and decoding settings are recorded; the grader, its kind, version and rubric hash are recorded and the grader is not the model under test; every task has exactly one graded result with an honest pass flag; and the published score, pass count, failures and near misses equal the recomputation.
- Does not prove: that the outputs were really produced by that model, that the grader graded well, that the training-hash list is complete, or that the benchmark measures anything useful. A SOUND verdict can carry a low score.
- Verify offline: `python3 verify_offline.py run.json --pinned-key POGikNM60jWUTaMMi5+r7w6WBAaiYm0dSKygmjh0ljw=`
