A military capability that depends on software can change materially without changing its programme name or its physical platform. A model can be retrained, its training data can be replaced, the processor or operating system that hosts it can be swapped during an obsolescence programme, and the circumstances in which commanders are authorised to use it can be widened by a change of doctrine. After each of these events the system may still look like the one that was tested, accepted or reviewed, while some of the evidence that supported those earlier decisions no longer describes its behaviour. The instruments with which European ministries answer this problem were, for the most part, not written for machine learning: the legal review of new weapons required by Article 36 of the First Additional Protocol, the qualification and assurance practices of their armament agencies, and the general terms on which they buy information technology. The question that matters for a European buyer, and for the suppliers that depend on European buyers, is therefore not whether “military AI” receives a single certificate that has to be renewed at intervals. It is which previous conclusion remains valid after a particular change, which authority is entitled to revisit it, what evidence is needed to do so, and whether the customer holds the contractual rights to obtain that evidence from the company that produced the change.