I certainly agree that there is an open question here. But I’m still tentatively optimistic — the whole reason we have a potential problem is because the model is computing in some way that generalizes perversely across states, and that gives us a potential handle. Like you said, a lookup table can’t cause trouble. And if B is using some heuristic to approximate A, it seems like that heuristic itself needs to generalize perversely, so I think we still have hope for being able to identify that.