Teams deploy whole systems but cite safety scores from base models. Fine-tunes drift safety in unpredictable directions, and many benchmarks are too noisy to detect the change. This post explains why safety scoring is not composable, and what MLCommons is doing to make it comparable.