STATE v1.0: Safety, Transparency, Accountability, Transferability, Ethics — a proposed buyer-side evaluation framework for clinical AI procurement

A hospital buying a clinical AI tool has one question: will this work here, on these patients, in this service? Regulatory clearance certifies that a tool may be sold in a jurisdiction. A composite procurement score certifies that its strengths outweigh its weaknesses on a weighted average. For clinical AI the second artefact is actively misleading, because the arithmetic lets a strong domain offset a failing one, and in a procurement decision that offset does not exist. A tool that misses two thirds of sepsis cases is not redeemed by good workflow integration. STATE v1.0 scores eight domains independently under five principles and applies a safety interlock: any domain below 30 of 100 forces Do Not Proceed regardless of the weighted total. Jurisdiction overlays for the GCC, Hong Kong, the UK and the US adjust the weights and localise the regulatory questions without changing the clinical ones. Two retrospective demonstrations, the Epic Sepsis Model at an August 2021 evidence vantage and Microsoft/Nuance DAX Copilot for a GCC buyer at a June 2026 vantage, were both decided by the interlock. The paper states its limits in full: it assesses documentary evidence at a stated vantage date, it has never run in a live procurement, its scoring involves judgement with measured inter-run variance, and its domains, weights and threshold are a reasoned first specification published for external critique and prospective testing.