There is a particular kind of problem that appears whenever a machine learns faster than the institution that is responsible for it. A conventional piece of mission software changes when someone decides to change it, and the decision leaves a trail: a requirement, a build, a test campaign, a release authority, a configuration record. A statistical model can change when new data arrives, and the change may leave no trace at all in the places where an armed service normally looks for one. The source code can be identical, the interfaces can be identical, the version string can be identical, and the behaviour of the system can be different. For a maritime helicopter whose principal job is to distinguish one underwater sound from another, that difference is the capability itself. The question this poses to a navy is not whether the model can be improved quickly. It is whether the organisation can still say, at any given moment, which model is flying, who authorised it, what evidence supported that authorisation, and how it would be withdrawn. For a European navy operating an American aircraft under an American release regime, the question doubles: even if those answers exist in Washington, it does not follow that they are available, or transferable, or exercisable in Copenhagen, Athens, Madrid or Bergen. That is the subject here — not the speed of retraining, but the location of control.