AgentSkillsCN

evidence-auditor

审核每项主张所依据的证据，并将存在的空白或疑虑记录在 `output/MISSING_EVIDENCE.md` 中。 **触发条件**：证据审核、证据缺失、未获支持的主张、审稿证据审计、证据缺口。 **适用场景**：在同行评审流程中，需逐条检查每一项主张的证据链，确认是否存在基准缺失，评估评审中的薄弱环节。 **若跳过**：如果尚未输入任何主张（例如，尚未创建 `output/CLAIMS.md`）。 **网络环境**：无需联网。 **行为准则**：仅记录“证据缺口/风险/下一步验证”，切勿替作者补充论述或引入新的主张。

SKILL.md
--- frontmatter
name: evidence-auditor
description: |
  Audit the evidence supporting each claim and write gaps/concerns into `output/MISSING_EVIDENCE.md`.
  **Trigger**: evidence audit, missing evidence, unsupported claims, 审稿证据审计, 证据缺口.
  **Use when**: peer review 流程中，需要逐条检查 claim 的证据链、缺 baseline、评测薄弱点。
  **Skip if**: 缺少 claims 输入（例如还没有 `output/CLAIMS.md`）。
  **Network**: none.
  **Guardrail**: 只写“缺口/风险/下一步验证”，不要替作者补写论述或引入新主张。

Evidence Auditor (peer review)

Goal: for each claim, either (a) point to the supporting evidence in the manuscript, or (b) write a concrete gap with an actionable fix.

Inputs

  • •output/CLAIMS.md

Outputs

  • •output/MISSING_EVIDENCE.md

Output format (recommended)

For each claim:

  • •Claim: copy the claim text
  • •Evidence present: what the paper provides (experiments/theory/citations)
  • •Gap / concern: what is missing or weak
  • •Minimal fix: the smallest additional evidence that would address the gap
  • •Severity: major | minor (optional)

Workflow

  1. •Iterate claims in output/CLAIMS.md.
  2. •For empirical claims, check:
    • •dataset/task definition is clear
    • •baselines are appropriate
    • •evaluation protocol is valid
    • •ablations/sensitivity analyses exist where needed
  3. •For conceptual claims, check:
    • •definitions are unambiguous
    • •assumptions are stated
    • •claims do not exceed what is argued
  4. •Write output/MISSING_EVIDENCE.md as a list of claim-by-claim entries.

Definition of Done

  • • Every claim from output/CLAIMS.md has an evidence note or a gap item.
  • • “Fix” items are actionable (what to add, not “more experiments”).

Troubleshooting

Issue: you cannot locate the evidence in the paper

Fix:

  • •Mark the claim as “evidence not locatable” and ask for a clearer source pointer (or re-extract claims with better pointers).

Issue: the audit starts proposing new claims

Fix:

  • •Stop; only critique what exists in output/CLAIMS.md and the manuscript.