Approved Once, Changed Daily

A single approval of an adaptive AI system certifies a system that will soon not exist, and without defined triggers for looking again the approval becomes a record whose main beneficiary is the organisation's own file rather than the people affected. This paper identifies four routes by which an approved system changes, being model updates, prompt and configuration changes, new data sources and new uses, and observes that only the first is routinely recorded, usually by the supplier. It proposes a nine-element continuous assurance model covering a versioned description, a change log, agreed reassessment triggers, performance and bias monitoring, an incident route open to affected people, renewed notice on a change of purpose, a named owner for supplier changes, and retirement criteria agreed before anyone has an interest in keeping the system running. Six triggers and four proportionality tiers keep the burden matched to the effect on people. The structure is borrowed from medical device regulation, where a system is approved together with an envelope of expected change. Four conditions for research ethics panels are added, since a study of a commercial tool faces the same problem.