When Evidence Updates Change Which Clinical QA Route May Win: State-Indexed Latency Evaluation

A latency percentile has no stable service meaning when an evidence update can change which physical routes still implement a clinical request. The route code and nominal runtime may remain fixed while a new contraindication makes its cached evidence stale, out of scope, or conflict-incomplete and removes that route from the eligible denominator. AdmiTrace-Q makes this membership transition executable. It compiles each query into four pre-execution certificate checks—provenance, temporal scope, conflict coverage, and freshness—ranks only routes certified against the current graph and cache watermarks, and records post-execution claim–evidence conformance without deleting failures. A fixed candidate ledger then separates measured opportunity, service-bucket exposure, and the deployed selector's response. On 847 clinician-authored questions, an unrestricted denominator lowers actual-oracle P95 from 3.24 s to 2.89 s; 98/847 (11.6%) questions contain a strictly faster ineligible route, 83/847 (9.8%) cross an actual-oracle service bucket, and 61/847 (7.2%) change the estimated-latency selector's decision. A disjoint 463-query public drug-label and guideline-update workload reproduces the three effects on 47 (10.2%), 36 (7.8%), and 29 (6.3%) questions. With all policies restricted to the same eligible set, adaptive selection reaches 3.24 s P95 and 4.87 s P99, versus 4.18 s and 6.92 s for OfflineGraph-Tuned, while a paired blinded audit improves completeness by +0.13 [+0.03, +0.23] and evidence F1 by +0.03 [+0.01, +0.06]. State-indexed eligibility prevents omitted evidence work from being rewarded as clinical QA speed and supplies a stable measurement target for systems serving mutable corpora.