Techniques to make model outputs more consistent and reliable, such as constraining output format or adding classification heads.