Commercial AI architectures operate without a deterministic epistemic trip state. When a generative model cannot retrieve grounded data, it continues operating: selecting statistically dominant tokens that satisfy grammar and structure regardless of factual accuracy. This is not a malfunction. It is the system operating as designed.

In any other safety-critical discipline - offshore engineering, aviation, pharmaceutical manufacturing - a system that continues operation under conditions of epistemic infeasibility is classified as defective by design. Generative AI is the only global information infrastructure operating without this classification applied.

The empirical baselines confirm the exposure. In legal practice, contra-factual error rates of 58-88%. In clinical medicine, failure rates exceeding 80% in early-stage diagnostic reasoning. These are the foreseeable consequences of an architecture with no refusal primitive and no commercial incentive to build one.

The work is structured around three interlocking components: a theoretical framework establishing the safety case; a physical implementation on air-gapped bare-metal demonstrating feasibility; and an audit protocol providing the legal and technical enforcement mechanism. The Phase 1 RAD commissioning run, logged above, demonstrates all three pipeline states: BLOCKED, FLAGGED with logged operator release, and CLEAN.