Obsidian Metadata
| type | concept |
| sources | Source - AI Evals FAQ, Nova Escola |
Reference Answers and Rubrics
- Reference answer: an example of a correct response.
- Rubric: criteria for judging a response (e.g., “follows the refund policy”).
Do you need one before annotating? No.
Writing a rubric before reviewing examples can get in the way:
- Reviewers get so focused on checking rubric items that they miss problems outside the rubric.
- Your idea of “good” will change as you review. That’s Criteria Drift, and it’s usually bigger than people expect.
- Example: a refund agent that escalates to a human exactly as policy says, but you only learn from reading interactions that the process frustrates users.
Treat initial expectations as a starting point you’ll revise.
Recommended sequence
- Do Error Analysis: open-ended notes, then group them to see what recurs.
- Then write the rubric from observed user and application behavior.
- Re-run error analysis periodically to keep the rubric current.
- When multiple annotators disagree, update the rubric with a definition, rule, or example for the disputed case and relabel (see Benevolent Dictator).
Cautionary tale: Nova Escola wrote a rubric before error analysis. Most criteria never showed up as failures, so the team labeled data it didn’t need.
Where reference answers help
- Reference-based checks against known correct values are cheap and reliable.
- CI regression sets for known workflows.
- RAG retrieval evals: query → relevant-chunk pairs.
- Anthropic recommends a reference solution per agent task to prove it’s solvable and that graders are configured correctly.
Production traffic usually has no reference outputs, which is why online monitoring leans on reference-free evaluators such as judges (see CI Evals vs Production Monitoring).
Related
Criteria Drift · Error Analysis · Eval Maintenance and Cadence