A Human-Centric Deterministic-Kernel Probabilistic-Shell Architecture with Invariants that Confine Clinical AI Decision Authority

Abstract

Medical artificial intelligence faces a structural tension: capability arises from probabilistic models, whereas safety requires guarantees those models cannot give. We implement and verify an architecture in which the probabilistic component holds no execution authority. A deterministic kernel is the sole output path and adjudicates every candidate through one fixed safety envelope. A human-centric core turns the patient's own account into decision weights, bounded so that no weight can widen the boundary. Verification runs along four lines: invariant, property-based, mutation, and adversarial. The formal lines passed but shared one blind spot, which a red team of 276,727 blind attacks from two vendors exposed: unenumerated input silently reaching the most permissive branch. We report this defect class, its remedy, and re-verification, including against the deployed system. Verification is architectural and does not establish clinical efficacy.