The Refusal Stack

Medical artificial intelligence is approaching a regulatory transition in which the burden of proof shifts from system performance to system accountability. The European Union Artificial Intelligence Act, the WHO guidance on large multi-modal models for health, and the United States Food and Drug Administration's frameworks for Software as a Medical Device each create or reinforce architectural expectations that benchmark-driven development does not address. This paper proposes that the operational unit of safety engineering for medical AI is the refusal — the system's documented, audited decision not to answer under specified conditions — and offers a five-class taxonomy for organising it. The paper describes one refusal class in detail, identifies architectural requirements for source-grounding, audit-trail immutability, and permission-first ingestion, and proposes a six-point technical due-diligence procedure. Implementations specific to one provider are not described.