Stefano Natangelo

Large language models are increasingly embedded in clinical decision support, but they are stateless. They preserve no epistemic commitments across contexts, identical cases can yield divergent outputs, and fluent generation can acquire de facto authority over what counts as evidence without any matching structure of responsibility. We argue that whether clinical AI can be trusted and held accountable depends on where epistemic authority sits and who governs it, not on how well the model performs or how fluently it generates. From this it follows that epistemic governance has to be externalized from the model. It also has to be constrained, because relocating authority into a second generative layer reproduces the original problem. We make the claim concrete with a non-agentic, constraint-based architecture, instantiated in oncology. Context classification, evidence authority, and truth-state assignment are settled by dedicated, inspectable layers before any generation occurs; the language model is confined to bounded extraction and summarization. The central move is to keep deontic authority, what is permitted and reimbursed, distinct from epistemic authority, what the evidence supports, so that a reimbursement or jurisdictional constraint is represented as an allocation decision rather than collapsed into an apparent clinical fact. Uncertainty, disagreement between authoritative sources, and clinician overrides are preserved as explicit, inspectable states. By settling authority before generation and keeping it inspectable, the architecture restores the structure of responsibility that fluent generation erodes: the exercise of authority becomes attributable, and the resulting recommendation can be contested rather than absorbed as plain clinical fact.