A Substrate-Layer Governance Architecture for Clinical LLMs: Regulatory-Grounded, Stakeholder-Aware, and Aligned with FDA, HIPAA, and ISO

Large language models are increasingly proposed for clinical decision support, documentation, and patient communication. The prevailing governance posture for these systems relies on input and output filtering — blocking unsafe prompts and screening unsafe responses. That posture is necessary but insufficient, because the risk in a clinical conversation is contextual rather than lexical: the same words are safe in one stakeholder context and unsafe in another, and a system can drift across a regulatory boundary without ever producing a flagged token. This paper proposes a substrate-layer governance architecture for clinical LLMs, organized in five layers — authority, orientation, lineage, evidence, and surface — and rooted in a single prior question: who is the user. Its risk categories are derived from the clinical regulatory regime (the FDA clinical-decision-support boundary, clinical accuracy and hallucination, HIPAA privacy, operational quality management under ISO, and clinical bias and equity) rather than from a generic harm taxonomy. The orientation layer, which monitors where a system is heading within a conversation, is presented as a thesis-layer control requiring empirical validation, not as a deployed guarantee. This is a proposed reference architecture, not a description of a validated or deployed system.