Provide a reusable structure for benchmark artifact files.
Use when documenting measured or estimated agent behavior.
- Date
- Scope
- Candidate stack
- Counting method
- Task text
- Baseline result
- Candidate result
- Quality scores
- Caveats
- Separate measured facts from interpretation.
- Label exact counts versus estimates.
- Include quality scoring.
- Preserve caveats.
- Avoid public benchmark claims unless the run is controlled.
A benchmark artifact under docs/benchmarks/.
- Honest scope.
- Reproducible enough for repo-internal use.
- Clear pass/fail decision.
- No overclaiming.