Clinical language model evaluation has concentrated on accuracy. The Coalition for Health AI organizes assurance around three principles: the other two are bias & fairness and safety & reliability. They are served by individual studies rather than by any framework.

This keynote introduces PacificMed and its 2026 edition: 70+ benchmarks across eight suites, derived from a standing register of 201 legal instruments maintained by practising lawyers and updated quarterly since 2024. Every case carries the provision it was written from and rolls up to one of twelve instrument families, each reported with its own confidence interval, so a failure points to the obligation a system does not currently meet. Benchmark sizes are derived from a stated precision target rather than from whatever a generation pipeline produced, cases are balanced jointly across deployment context and harm level, and every case is reviewed against six criteria by a clinician, privacy or security practitioner.

PacificMed is contributed to the open-source MedHELM project as private scenarios, free for academic research under a data use agreement. It covers four categories with no counterpart in any existing medical safety benchmark, five security constructs a clinical taxonomy does not contain, and the 2025 to 2026 state statutes on AI therapy, medical licensure and patient communication.

This session describes an evaluation framework and the legal instruments its cases are derived from. It is not legal advice and not a compliance certification.