Bounding Repetition in Clinical AI Safety Arguments from Fail-Operational to Dynamic Trajectory Governance

Abstract

Safety arguments for clinical artificial intelligence are stated per encounter: a proposal is admitted or refused against the patient's state at one moment. We show that such a formalism cannot express repetition. In a released deterministic kernel the treatment history is retained but consumed only through a set projection, and the dose is discarded when an action is applied; a candidate admissible once is admissible without bound, and no predicate changes value. Enumerating encounter sequences over the same rule set, 167,165 admissible trajectories at length eight collapse onto 59 states the kernel can distinguish, under a ceiling of 384 that does not grow with follow-up. We state invariants bounding cumulative exposure, unchanging repetition, and the scope and span of physician override. Across 28,846 adversarial trials from four models no override exceeded its scope or span; instead 12 per cent of admitted requests claimed that authority in prose.