Why LLM Evals Are Different

Classical software has a spec: inputs → expected outputs, verified by deterministic tests. An LLM's spec is the prompt, its output is open-ended text, and the same input legitimately produces a hundred different correct answers.

That changes what 'testing' means. There's no assertion that captures 'a helpful summary' — correctness is a judgment call, not a type check. And the nondeterminism isn't a bug to fix: temperature, sampling, and model updates mean yesterday's pass can be today's fail for reasons that have nothing to do with your change.