Governing AI Evidence in Clinical Trials: Protocol Interpretation, Decision Stability, and Predictive Validity

Protocol-dependent AI evidence can agree across evaluated interpretations while omitting supported alternatives or relying on an inaccurate model. This article develops a claim-specific governance framework that separates interpretation coverage, conditional decision stability, and predictive validity. Finite-domain results show why agreement and low clarification cost cannot certify coverage and why a raw question-selection heuristic can waste arbitrarily many questions on irrelevant detail. Existing computational and exploratory clinical illustrations are retained as prior evidence rather than new validation. The journal extension contributes an executable specification-integrity benchmark with relational constraints, source precedence, amendments, conflicts, and delegated choices, together with a versioned claim-assessment and dependency-invalidation demonstrator. Across 384 held-out complete synthetic references, marginal-product outputs preserved every reference but added unsupported worlds in 315 cases. A separate implementation matched all 1,152 serialized reference records and detected all 1,344 prespecified faults. Broader structural fixtures exposed missed stale claims when dependencies or changes were untracked. The evaluations test finite software semantics and constructed fault detection, not natural-language interpretation, clinical validity, or human review effectiveness. A proposed accountability process assigns different responses to omitted readings, decision-changing uncertainty, predictive mismatch, and changed evidence. The central policy implication is to bind every reported conclusion to its evaluated domain, source authority, model, intended use, and reassessment conditions. A complete synthetic record can make those dependencies inspectable; it cannot supply missing scientific evidence or legitimate an unauthorized choice. Extended version of 'AI Policy for Clinical Trial Evidence: Governing Protocol Ambiguity in Simulation,' accepted at the World Models for High-Stakes Health (WMHS) workshop at NeurIPS 2026 (non-archival). 25 pages, 4 figures.