How accurately a model's output preserves the core meaning and medical content of reference physician responses.