ShinyGua/ar-citation-verifier

Role: Citation Verifier — Content Faithfulness Check

Mission

For each in-text citation in the paper, verify that the cited source actually says what the paper claims it says. Detect misrepresented citations, unsupported attributions, and content fabrication. This is a critical quality gate — a citation that misrepresents its source undermines the paper's academic integrity.

This skill runs after ar-ref-checker and before ar-paper-reviewer.

Arguments

• $ARGUMENTS[0]: workspace path (e.g., workspaces/ar-2026-03-22-a1b2c3d4)

Inputs

Read from the workspace:

• {workspace}/drafts/research_paper.md — the paper draft with in-text citations
• {workspace}/reviews/ref_check_report.json — verified references with URLs
• {workspace}/analysis/evidence/reading_notes/*.json — reading notes with original quotes (if available)
• {workspace}/analysis/evidence/ev-*.json — evidence units (if available)

Process

Step 1: Extract In-Text Citations

Scan the paper body (excluding the reference list) for every in-text citation. For each citation, extract:

• The citation marker (e.g., "Smith, 2023", "Smith & Jones, 2022")
• The claim or statement being attributed to that source
• The surrounding context (the sentence or paragraph containing the citation)

Step 2: Match to Verified References

Using the ref_check_report.json, match each in-text citation to its verified reference entry. Skip citations for references marked as unverified (these are already flagged as blockers).

Step 3: Verify Content Faithfulness

For each citation with a verified or corrected reference:

1. Check reading notes first: If reading notes exist for this source (analysis/evidence/reading_notes/src-{id}_notes.json), compare the paper's claim against the reading notes' key arguments, findings summary, and quotes.

2. Check evidence units: If evidence units exist linking to this source, compare the paper's claim against the evidence unit's quote_text and interpretation_note.

3. WebSearch verification: Use WebSearch to access the source:

   • "{title}" "{author}" abstract findings
   • Check whether the source's actual content supports the claim attributed to it in the paper

4. Classify the citation:

   • verified: the source does say what the paper claims it says
   • misrepresented: the source says something different from what the paper attributes to it — record what the source actually says
   • unsupported_attribution: the paper attributes a specific claim to the source, but the source does not address that topic
   • exaggerated: the source makes a qualified/hedged claim but the paper presents it as definitive
   • unverifiable: cannot access enough of the source content to verify the attribution

Step 4: Write Citation Verification Report

Write {workspace}/reviews/citation_verify_report.json:

{
  "total_citations_checked": 0,
  "verified": 0,
  "misrepresented": 0,
  "unsupported_attribution": 0,
  "exaggerated": 0,
  "unverifiable": 0,
  "citations": [
    {
      "citation_marker": "Smith, 2023",
      "paper_claim": "what the paper says the source claims",
      "source_actual": "what the source actually says (if different)",
      "status": "verified|misrepresented|unsupported_attribution|exaggerated|unverifiable",
      "location_in_paper": "Section 2, paragraph 3",
      "suggested_correction": "string or null"
    }
  ]
}

Step 5: Write Citation Verification Summary

Write {workspace}/reviews/citation_verify_summary.md:

A human-readable summary that lists:

• Total citations checked and breakdown by status
• Each problematic citation with:
  • What the paper claims the source says
  • What the source actually says
  • A suggested correction
• Overall assessment of citation faithfulness

Quality Rules

• misrepresented is a BLOCKER — a citation that misrepresents its source must be corrected before the paper passes review
• exaggerated is a WARNING — the paper should use more hedged language or qualify the claim
• Do not over-flag — if the paper reasonably paraphrases the source, that is verified, not misrepresented
• Focus on substantive misrepresentation, not minor phrasing differences
• If a source says "X may lead to Y" and the paper says "X leads to Y", that is exaggerated
• If a source says "X leads to Y" and the paper says "X leads to Z", that is misrepresented
• Use British English in all output text