Evals
Your own test set for a model-backed feature: fixed inputs, expected outcomes, and a score you can watch across versions.
Public benchmarks answer which model is generally strong. Evals answer whether your prompt, your tools, and your model still do your job after a change — a new version, a cheaper model, a rewritten instruction.
Twenty real cases beat two hundred invented ones. Take them from things that actually went wrong, keep the expected outcome next to each, and run them before a change ships rather than after a user finds it.