RAGF-H: A Domain-Specific Regulatory Governance Framework for Agentic AI in EU-Regulated Healthcare

We present RAGF-H, an extension of the Reflexio Agentic Governance Framework (RAGF) addressing three architecturally distinct regulatory compliance problems that arise when autonomous AI agents operate in EU-regulated clinical environments. The delegation of clinical acts is legally constrained by professional licensure — not authentication — making the active registration status of the responsible practitioner a runtime precondition of action validity. Access to health data constitutes a regulated processing event under GDPR Art. 9 independent of any subsequent agent action, breaking the action-first evaluation flow that financial-domain governance frameworks assume. And four overlapping regulatory regimes — EU AI Act Annex III, MDR/IVDR Rule 11, GDPR Art. 9, and Spanish LOPS/LBAP — produce simultaneously active obligations whose precedence cannot be modelled as a total order. RAGF-H introduces three contributions: (1) a Professional Credential Validator (PCV) that operates as a pre-gate performing active licensure verification against national professional registries (CGCOM, CGCOF, equivalents), with a documented three-mode graceful degradation protocol bounding behaviour during registry unavailability; (2) a Data Category Pre-Gate (DCPG) that triggers GDPR Art. 9 evaluation on data access events, not only on action proposals; and (3) a Dynamic Precedence Graph (DPG) that resolves multi-regulation conflicts through conditional dominance relations between constraints, escalating irreducible conflicts to human legal review with full audit-trail traceability. The framework preserves the RAGF invariants of deterministic fail-closed enforcement, HMAC-chained audit trails, and sub-millisecond hot-path latency, while adding full verdict traceability to specific regulatory articles. A five-use-case experimental validation protocol is specified — covering triage assessment, prescription support, clinical-trial enrolment, MDR adverse-event vigilance, and discharge coordination across AMM levels L2–L4 — against three baselines (no governance, probabilistic guardrail, unextended RAGF). This is a working draft (v0.7). Experimental results in Section 4.3 are stated as targets and hypotheses pending empirical execution against the reference implementation; they are not validated measurements. The version is suitable for preprint deposit; a future version with populated empirical results will accompany conference submission. Reference implementation: github.com/cronocom/ragf. This work extends the base RAGF framework (Zenodo DOI 10.5281/zenodo.20285406), whose Spanish healthcare deployment of 1,893 medication-safety actions established operational viability that RAGF-H now formalises at the regulatory-compliance layer.