Validate correctness before counting performance gains

A fast error response is not a performance improvement. Use the same warmup, count, and concurrency protocol as the baseline; reject samples whose status or content contract fails; and exercise representative outcome-specific branches. Correlate structured outcome and timing evidence, reproduce the smallest failure under the exact runtime configuration, and quantify branch reach before rollout judgment.

Diagnosis does not prove a proposed fix until it is implemented and the same live protocol passes.