A model can predict what might happen next and still not know what to do.

That gap is where I keep circling: action-relevant state, context faithfulness, and whether a world model is honest about what it does and does not know.