Model performance measured on the same fixed distribution of prompts, independent of which model is being evaluated.