A model comparison needs one unchanged prompt and one fixed scoring note. Otherwise a better answer may just match a better question.