Evaluation·Also: Eval set / Golden dataset
Golden set
A curated collection of representative inputs with known expected behaviour, used as the regression backbone of an AI evaluation harness.
A golden set is the curated backbone of an evaluation harness. It is a version-controlled collection of inputs - typically 100 to 500 - with expected outputs or expected properties, used to catch regressions before customers do.
What goes in one
A good golden set is:
- Representative. If 20% of production traffic is non-English, the golden set should be roughly 20% non-English.
- Inclusive of the long tail. Edge cases, adversarial phrasings, low- confidence domains. The things that break in production, not the things that worked in the demo.
- Labelled with expected behaviour. Exact output where ground truth is clean; a rubric where it isn't.
- Maintained. Production sampling feeds reviewed examples back into the set so it stays representative as traffic evolves.
Common failure modes
- A golden set of 20. Too small to detect regression statistically.
- Only happy-path inputs. The set passes; production fails on the long tail.
- Stale. No update in six months. Traffic has evolved; the set hasn't.
Why it's a governance artifact
In regulated deployments, the golden set is one of the first things an auditor asks for. It is the evidence that the system is measurably behaving as designed, not just claimed to be.