Clinical training has, for over a century, been organised around three institutional commitments that medicine and the surgical profession have considered non-negotiable: scenarios must be reproducible across cohorts, the trajectory of a trainee’s response must be traceable to an examinable record, and the pedagogical artefact must be inspectable by the body responsible for accreditation. The arrival of artificial intelligence in clinical pedagogy does not introduce these requirements. It either preserves them or degrades them. This paper proposes that the operational unit of safety engineering for AI-mediated clinical training is the audited scenario: a versioned, multi-phase pedagogical artefact whose every learner interaction is traceable, whose evaluation is decomposable into named pedagogical objectives, and whose institutional aggregation surfaces cohort-level patterns of error without exposing individual learners. It identifies four architectural commitments specific to AI-mediated pedagogy: reproducibility of the scenario, traceability of the learner trajectory, decomposability of evaluation into named pedagogical objectives, and institutional aggregation without individual exposure. One of these commitments — decomposability of evaluation — is described in operational detail through four named components: rubric atomicity, weighting transparency, score-to-criterion traceability, and rubric versioning discipline. The paper also proposes a six-point pedagogical due-diligence procedure for institutions evaluating AI-mediated training systems, and three evaluation axes: scenario reproducibility, trajectory completeness, and dashboard fidelity. The architecture inherits four foundational commitments from companion technical work on source-grounded retrieval, permission-first governance, refusal as a first-class behaviour, and cryptographically chained audit trails. No source code, deployment architecture, or proprietary implementation detail is disclosed. Part of a CLINETHIX preprint series on defendable clinical AI infrastructure, 2026. Related works: The Refusal Stack, doi:10.5281/zenodo.20257894; Sovereignty by Design, doi:10.5281/zenodo.20258141; Companion — Ambient Clinical Assist Under Human Oversight by Design, CLINETHIX preprint series, 2026.
Fatima Azzahra MASTARI (Mon,) studied this question.