An evaluation dataset should include ordinary tasks, edge cases, prior failures, and adversarial examples from the real distribution. Its labels and rubrics need versioning because the definition of acceptable behavior changes.