Files
TWBG_Materials/文档润色流和知识库构建流/claude-scholar-upstream/skills/nature-response/tests/evaluation-summary.md
2026-05-30 16:22:29 +08:00

44 lines
2.2 KiB
Markdown

# Evaluation summary
`nature-response` is evaluated with synthetic Markdown fixtures. These tests are not executable
unit tests; they are behavior contracts for manual and agent review.
## Status rationale
Recommended status: `Beta`.
Rationale:
- The core rules are defined in `SKILL.md` and modular references.
- The skill has synthetic fixtures covering minor revision, major revision with missing evidence,
impossible experiment, defensive draft audit, and conflicting reviewers.
- Each fixture includes expected behavior, forbidden behavior, and pass/fail criteria.
- The examples show expected output shape without using real confidential reviewer comments.
- The skill has not yet been validated on real anonymized revision packages, so `Stable` would be premature.
## Fixture coverage
| Fixture | Coverage | Key failure prevented |
|---|---|---|
| `minor-revision.md` | stable IDs, minor comments, missing citation metadata | fabricated citation or line numbers |
| `major-revision-missing-evidence.md` | validation request, statistical details, missing evidence | invented results or p-values |
| `impossible-experiment.md` | out-of-scope longitudinal evidence | time/funding excuse or fabricated survival data |
| `defensive-draft-audit.md` | hostile draft language, vague compliance | accusatory reviewer language |
| `conflicting-reviewers.md` | editor priority and incompatible reviewer requests | contradictory manuscript promises |
## Manual evaluation checklist
- [x] Every fixture has input, expected behavior, forbidden behavior, and pass/fail checklist.
- [x] No fixture uses real reviewer comments.
- [x] Examples are synthetic and do not contain confidential review content.
- [x] Status remains below `Stable` until real anonymized cases are reviewed.
## Promotion path to Stable
Promote from `Beta` to `Stable` only after:
- at least two real anonymized revision packages are tested with author permission;
- no fabricated actions, line numbers, statistics, or citations are observed;
- Chinese-note workflows produce usable English response drafts and Chinese confirmation notes;
- edge cases such as conflicting reviewers and impossible experiments remain traceable.