A benchmark harness default is not model-performance evidence

A benchmark harness default is configuration, not evidence that the default model performs best. Emergency provider substitutions should use recorded, like-for-like task results from the same population and scoring contract. If labels were not independently adjudicated, call the result compatibility with the incumbent rather than correctness.

Applicability and limits: this applies to structured extraction and OCR replacements. A higher compatibility count does not establish truth, production latency, cost, or live rate-limit behavior. A model with no executed result must remain unmeasured even if the repository names it as the default.