“After hallucination-reduction training, the model did not become overly cautious or refuse to answer questions, and the number of claims it made in long-form completions did not decrease.”