citation-claim-audit

Citation Claim Audit

Audit claims before they become paper text, README claims, or wiki synthesis.

Read First

• references/repository-contract.md
• references/claim-audit.md
• references/output-contracts.md

Workflow

1. Extract atomic claims.
2. Classify each claim: bibliographic, descriptive, methodological, comparative, causal, interpretive, or result.
3. Locate cited evidence and, when needed, the raw PDF or experiment artifact.
4. Check the exact support span: page, section, table, figure, equation, commit, run ID, dataset row, or source note.
5. Assign a verdict.
6. Record unsupported, contradicted, stale, or over-broad claims.
7. Propose minimal edits that preserve the supported meaning.

Claim Scope

Audit all numerical, factual, comparative, causal, novelty, SOTA, and methodology claims. For long documents, sample only when the user accepts the sampling plan; final manuscript checks should cover every acceptance-critical claim.

For reports/paper/sota-survey.tex, audit every durable survey claim that summarizes the field, compares methods, states novelty, reports metrics, or derives implications for the current research idea. Do not sample final SOTA survey claims unless the user explicitly accepts a sampled audit.

Audit Table

Use this structure in analysis_outputs/claim-audit.md or CSV:

claim_id	claim	class	cited_source	evidence_path	support_span	verdict	fix

Verdict Discipline

• supported: exact or clearly entailed by evidence.
• partially-supported: true only in narrower scope.
• unsupported: no evidence found.
• contradicted: evidence conflicts.
• wrong-source: source does not support the claim.
• needs-human: inaccessible or needs domain judgment.
• stale: source was superseded or the result changed after the cited evidence.

Fix Discipline

Prefer the smallest safe change:

• add missing citation only when the source was verified
• weaken wording to match source scope
• split one broad claim into narrower supported claims
• move speculation to limitations/future work
• remove the claim when support is absent
• add an experiment or analysis task when the claim belongs to project data

Do Not

• Fabricate missing references.
• Treat citation count as truth.
• Trust PDF-to-Markdown for exact numbers without checking the source.
• Collapse "unsupported" into "false".
• Strengthen claims during editing.
• Accept a bibliography match as proof that the cited passage supports the claim.