> When a model exhibits hallucination, often providing more context and evidence will dispel it,

I usually have the opposite experience. One a model goes off the rails it becomes harder and harder to steer and after a few corrective prompts they stop working and it’s time for a new context.