claims-extractor（索赔提取器）

Claims Extractor (paper review)

Goal: turn a manuscript into an auditable list of claims that downstream skills can check.

Inputs

Required:

• output/PAPER.md (or equivalent plain-text manuscript)

Optional:

• DECISIONS.md (review scope or constraints)

Outputs

• output/CLAIMS.md

Output format (recommended)

For each claim, include at minimum:

• Claim: one sentence
• Type: empirical | conceptual
• Scope: what the claim applies to / what it does not apply to
• Source: a locatable pointer into output/PAPER.md (section + page/figure/table + a short quote)

Workflow

1. If DECISIONS.md exists, apply any review scope/format constraints.
2. Read the manuscript (output/PAPER.md) end-to-end (at least abstract + intro + method + experiments + limitations).
3. Extract:

- primary contributions (what is new) - key claims (what is asserted) - assumptions (what must be true for claims to hold)

1. Normalize each item into one sentence.
2. Attach a source pointer for every item.
3. Split into two sections:

- Empirical claims (must be backed by experiments/data) - Conceptual claims (must be backed by argument/definition)

Definition of Done

• output/CLAIMS.md exists.
• Every claim has a source pointer that can be located in output/PAPER.md.
• Empirical vs conceptual claims are clearly separated.

Troubleshooting

Issue: the paper is only a PDF or HTML

Fix:

• Convert/extract it into a plain-text output/PAPER.md first (even rough extraction is OK), then run claim extraction.

Issue: claims are vague (“significant”, “better”, “state-of-the-art”)

Fix:

• Rewrite each claim to include the measurable dimension (metric/dataset/baseline) or mark it as “underspecified” with a note.