An adaptive interface can look impressive right up to the moment a user asks, “Where did the export button go?”
That question creates an uncomfortable tension for developers. The model may be capable of producing a plausible layout, but plausibility is not the same as operational safety. Support cannot investigate a screen that no longer exists, and developers cannot improve a system when the only evidence is a screenshot of an ephemeral UI.
The answer is not to preserve every pixel or require human approval for every spacing change. It is to treat generated UI as a versioned proposal with three properties:
It can express only approved components and actions.
Every rendered version can be identified and replayed.
