The Messing paper's most consequential finding is not that evaluation scores are noisy — that has been suspected — but that the noise is exploitable in a specific direction . When variance goes unreported, the entity best positioned to probe the measurement surface is the model developer, not the evaluator. Labs running repeated evals against the same benchmark can identify which prompt phrasings, judge configurations, and temperature settings favor their model's outputs, then ship the version that scores highest on those conditions. The benchmark community has no mechanism to detect…