What belongs in it
A useful set contains ordinary examples, edge cases, known failures, ambiguous requests, and cases where the correct behavior is refusal or escalation. The examples should come from the real job, not only from a developer’s imagination.
Why it is called golden
- The expected answer or label has been reviewed.
- The input is stable enough to compare across runs.
- The set is protected from casual edits.
- Changes are versioned and explained.
- It is small enough to run frequently but broad enough to expose risk.
What it cannot tell you
A golden set is not a guarantee of production quality. It can miss new user behavior, new data, and rare failures. Use it as a regression alarm, then add real reviewed failures back into the set over time.