# CAIN-A75-058 Framework Evaluation Review

**Verdict:** VERIFIED

> Recomputes a repeated-run framework comparison: frozen independent harness, pinned model and tools, full coverage, mean and variance per framework, ranking, raw-results hash, and whether the leader is outside the noise. It does not re-run anything.

| Study | Verdict | Ranking (mean ± sd) | Leader margin vs noise | Failed checks |
|---|---|---|---|---|
| 0 study:agent-framework-comparison-sample | SOUND | fw:alpha 0.9167±0.1443, fw:beta 0.6667±0.1443, fw:gamma 0.3333±0.1443 | 0.25 vs 0.1666 |  |

## What this proves and what it does not
- Proves: the harness hash over the tasks and rubric is recomputed and it was frozen by a named human who is not a vendor of a compared framework; every run used the frozen harness, the pinned model, temperature and tool versions, after the freeze; every framework covered every task in every repetition exactly once; the published means, standard deviations, run counts, ranking and raw-results hash equal the recomputation; and a leader is called separated only when its margin exceeds twice the larger standard error.
- Does not prove: that the runs really happened as recorded, that the tasks are representative, or that two standard errors is the right bar for your decision. It is a study-record review.
- Verify offline: `python3 verify_offline.py run.json --pinned-key 6v355Y7MDhavg3kvnVm2Q2Ozu5ny1tPKolVaq7/r/c0=`
