Agent Skills: Claim–Evidence Matrix

Claim–Evidence Matrix

Triggers & routing

• Trigger: claim–evidence matrix, evidence mapping, 证据矩阵, 主张-证据对齐.
• Use when: 写 prose 之前需要把每个小节的可检验主张与证据来源显式化（outline + paper notes 已就绪）。
• Skip if: 缺少 outline/outline.yml 或 papers/paper_notes.jsonl。
• Network: none.
• Guardrail: bullets-only（NO PROSE）；每个 claim 至少 2 个证据来源（或显式说明例外）。

Make the survey’s claims explicit and auditable before writing prose.

This should stay bullets-only (NO PROSE). The goal is to make later writing easy and to prevent “template prose” from sneaking in.

Inputs

• outline/outline.yml
• papers/paper_notes.jsonl
• Optional: outline/mapping.tsv

Output

• outline/claim_evidence_matrix.md

Workflow (heuristic)

Uses: outline/outline.yml, outline/mapping.tsv.

1. For each subsection, write 1–3 claims that are:
   • specific (mechanism / assumption / empirical finding)
   • falsifiable (“X reduces tool errors under Y evaluation”, not “X is important”)
2. For each claim, list ≥2 evidence sources:
   • prefer different styles of evidence (method paper + eval/benchmark paper, or two competing approaches)
3. Keep it tight: claim → evidence → (optional) caveat/limitations.
4. If evidence is weak or only abstract-level, say so explicitly (don’t overclaim).
5. If bibkey exists in papers/paper_notes.jsonl, include [@BibKey] next to evidence items to make later prose/LaTeX conversion smoother.

Quality checklist

• [ ] Every subsection has ≥1 claim.
• [ ] Each claim lists ≥2 evidence sources (or an explicit exception).
• [ ] Claims are not copy-pasted templates (avoid “围绕…总结…” boilerplate).

Helper script (optional)

Quick Start

• uv run python .codex/skills/claim-evidence-matrix/scripts/run.py --help
• uv run python .codex/skills/claim-evidence-matrix/scripts/run.py --workspace <workspace>

All Options

• See --help (this helper is intentionally minimal)

Examples

• Generate a first-pass matrix, then refine manually:
  • Run the helper once, then refine outline/claim_evidence_matrix.md by tightening claims and adding caveats when evidence is abstract-level.

Notes

• The helper generates a baseline matrix (claims + evidence) and never overwrites non-placeholder work; in pipeline.py --strict it will be blocked only if placeholder markers remain.

Troubleshooting

Issue: claims are generic or read like outline boilerplate

Fix:

• Tighten each claim to a falsifiable statement and add an explicit caveat if evidence is abstract-only.

Issue: you cannot add [@BibKey] because keys are missing

Fix:

• Run citation-verifier to generate citations/ref.bib, then use the produced keys in the matrix.