| name | evidence-extraction |
| description | Claim-evidence extraction for Idea2Repo research artifacts. Use when building or reviewing claim matrices, evidence rows, paper notes, or downstream scoring inputs that must cite PDF page/quote/chunk ids. |
Evidence Extraction
Only promote claims that are backed by verified PDF evidence.
Workflow
- Read PDF chunks and manifest records.
- Extract claims about problem, method, baselines, datasets, metrics, limitations, and relation to the current idea.
- Keep each row tied to
paper_id, page, quote, chunk_id, status, and confidence.
- Write
docs/reference/claim_evidence_matrix.csv and paper note markdown.
Guardrails
- A quote without a page or chunk id is not verified evidence.
- Do not upgrade candidate metadata into evidence.
- Downstream related work, novelty, and scoring must use verified evidence rows only.