Evals
What it means
What are evals?
Evals are repeatable tests that score an AI system's outputs against a fixed set of examples and criteria. A team runs them before release and again whenever the prompt, data or model changes.
The comparison that matters is to an ordinary test suite. There is one difference: the system under test can give a different answer each run. A typical eval is a set of inputs with known good answers or a scoring rule. It runs automatically whenever the prompt, the model or the surrounding code changes. The team tracks the pass rate over time rather than a single pass or fail.
Engineering teams build them because "it worked when I tried it a few times" does not survive a prompt change six months later, once nobody remembers which cases mattered. Security reviewers and boards ask about them because an eval suite is the concrete answer to "how do you know it still works." Without an eval suite, that question has no answer better than trust.
In practice, evals change how a team ships. A prompt or model change runs through the same suite every time. A drop in the pass rate blocks a release, the same way a failing test would.
Published · Updated
In practice
Why it comes up
A security reviewer or a technical board member asks the same question they would of any release process. What stops a bad change from reaching customers? For a normal codebase, the answer is a test suite. For an AI feature, the honest answer is either an eval suite or “we check by hand.” That second answer does not hold up once a product has more than a couple of prompts.
What to do about it
Write down twenty to fifty real cases the system must handle correctly. Draw them from actual support tickets or usage, not invented examples, and include a few you know are hard. Score each case automatically where you can, and by a human reviewer where you cannot. Track the pass rate over time rather than treating one run as final. Run the suite before every prompt or model change ships.