EviRoute: Planning Clinical Audits with Finite-Cohort Certificates

A clinical AI alarm creates an immediate planning problem: which team should investigate, and what evidence should it collect? Changes in model error, reference procedures, and outcome confirmation can produce the same observed alarm. EviRoute converts this ambiguity into a budgeted audit. Three randomized review arms complete a time-by-process risk matrix; a confidence-calibrated planner selects the least-cost qualifying allocation; and exact finite-cohort sets support owner-specific investigation routes. The construction joins evidence acquisition and action selection through a shared statistical contract, providing a decision component for agent-assisted clinical research. In 112,000 computational replications at an operational certificate error level of 0.015, exact inversion produced three issued noncoverage events and seven withheld certificates. A matched Wald construction produced six and 1,337, respectively. Across four near-boundary planning profiles and 100 independent planning replications, 397 of 400 requests were certified at 1,000 planning trials per candidate and scenario; every issued design passed independent evaluation. These results connect finite-cohort uncertainty to an executable plan for the next investigation.