Toward an Evaluation Science for Generative AI Systems