Guideline authorities (WHO, NICE, HAS, NEJM, BMJ, JAMA) currently distribute clinicalguidance as PDFs and web pages. General-purpose large language models ingest, paraphrase,and re-emit this guidance at a conversational scale — stripped of provenance, version, andjurisdiction. This paper names that failure mode "provenance erosion" and argues thatclinical guidance now needs its own distribution rail: a signed, versioned, machine-A queryable channel through which authoritative content can be retrieved by AI systemsunder the issuing body's control. The paper is addressed to regulators (EU AI Act competent authorities, FDA DigitalHealth Center of Excellence, MHRA, Health Canada) and to the guideline authoritiesthemselves. It maps the gap between current AI-Act obligations (Articles 9-15, 50;Annex III §5) and the operational reality of clinical AI deployments, and proposesfive concrete instruments — guideline sovereignty registries, attestation feeds,binding refusal taxonomies, audit-grade trails, and bilateral conformity assessment —that turn high-level governance principles into deployable infrastructure. The companion technical paper, "The Refusal Stack" (Mastari, 2026,doi:10.5281/zenodo.20257894), provides the engineering primitives — refusal classes,retrieval-bound generation, prompt-injection containment — that makes the policyThe framework in this paper enforceable. Part of a three-paper series on defendable medical AI infrastructure published byCLINETHIX (2026).
Fatima Azzahra MASTARI (2026) studied this question.