Citation Verification

Never let a fabricated, misattributed, or metadata-wrong citation reach the user.

Scope

Covers academic citations only: PMIDs, DOIs, PMC IDs, arXiv IDs, NCT numbers, paper titles, author names, journal references, author-year citations, and bibliographies.

Does NOT cover non-academic verifiable specifics (jurisdictions, fees, addresses, drug doses not tied to a cited paper, sports/finance specifics, officeholders, policy dates) — fabrication-audit handles those; see Integration below.

When to Trigger

DEFAULT: ON for any output containing references. This is a background verification layer, not a standalone task.

Always active for: manuscripts, abstracts, grant narratives, aims pages; literature reviews, digests, article summaries; clinical consults citing specific studies as evidence; rebuttal letters; teaching or board-review content citing sources; emails/memos referencing specific papers; any response naming a study (author, year, journal, or title).

Skip for: general physiology explanations with no specific study cited; casual conversation; code/data tasks with no references; user says "don't worry about verifying references"; non-academic specifics (route to fabrication-audit).

Also trigger on explicit requests: "verify these references," "check my citations," "audit my bibliography," "are these real papers?," "did I cite this correctly?"

Core Principle

Never generate a citation from memory alone; every reference is verified through search before inclusion. Failure after the attempt limits below → [CITATION NEEDED], told to the user.

Verification Workflow

During Writing (Inline Verification)

1. SEARCH → PubMed MCP (title + first author); non-PubMed-indexed work falls back to OpenAlex / Semantic Scholar, then web search.
2. CONFIRM → the paper exists and metadata matches.
3. PMID/DOI → look up via get_article_metadata, never from memory.
4. CHECK → the cited claim actually appears in that paper.
5. FORMAT/CITE → use the verified metadata (author, year, journal, PMID/DOI) in the output.

This loop runs for every citation.

PMID Verification

PMIDs are the highest-risk hallucination field: other fields can be correct while the PMID resolves to an unrelated paper (references/history.md). Required workflow for every PMID:

1. Search PubMed MCP by title + first author.
2. Extract the PMID from the result, not from memory.
3. Call get_article_metadata with that PMID to confirm the returned title/authors match.
4. A different returned paper means the PMID is wrong; search by title instead.
5. No PMID confirmed after 2 attempts → omit it and use the DOI only, or mark [PMID NEEDED].

Never generate a PMID from memory, and never claim one is "verified" without the tool call.

DOIs

Verify a DOI resolves to the correct paper via web search or PubMed; more reliable than PMIDs but still warrants a check for new or recent papers.

arXiv IDs

Fetch https://arxiv.org/abs/<id> (must return 200) and confirm title + first author match. A 404 means the ID is wrong: search by title (OpenAlex/Semantic Scholar) and correct it, or mark [CITATION NEEDED]. If later published, prefer the published DOI/PMID and keep the arXiv ID only if it still resolves.

Retraction / Erratum Check

A citation can resolve with valid metadata yet be retracted or corrected; citing it is a hard error. Screen by parsing CommentsCorrections for RefType=RetractionIn, ErratumIn, or ExpressionOfConcernIn, not by publication type (misses corrections on an otherwise-normal record, references/history.md). A publisher or Retraction Watch banner is supplementary, never primary. Flag inline: [RETRACTED — <topic>; verify before citing] or [ERRATUM — <what changed>; verify before citing].

Identifier Extraction

Use [0-9]{5,9} for PMIDs, not [0-9]{7,8}, which silently drops shorter IDs (references/history.md). State the extraction pattern alongside any verification count so the denominator can be audited.

Metadata That Must Match

Claim Verification

When citing a paper for a specific claim (e.g., "mortality was 30%"): search the abstract or full text, confirm the finding appears, note if it is a secondary/subgroup result, and if unconfirmable write "Author et al. reported [claim] (exact finding not confirmed from abstract; verify in full text)."

Verification by Context

• Manuscripts/grants: verify every bibliography reference during drafting, not post-hoc; flag failures with [CITATION NEEDED]; literature review sections need at least title + first author + year + journal.
• Clinical consults: verify the paper and finding; landmark trials (ARDSNet, PROSEVA, TTM2) must be attributed to the correct trial and outcome; a dose tied to a cited paper is verified here, free-standing goes to fabrication-audit.
• Literature digests: confirm every link and title exactly (capitalization, subtitles); Medline-style journal abbreviations.
• Rebuttal letters: full verification for new references; re-cited papers must be confirmed in the bibliography with the finding intact.

Handling Verification Failures

Paper not found: try alternate search queries, then the step-1 fallback order (PubMed MCP → OpenAlex → Semantic Scholar → web search); confirm first author + year on any title-search hit, since relevance ranking can surface a same-title decoy; still not found after 2-3 attempts → [CITATION NEEDED].

PMID not found or mismatched: use the correct PMID if found under a different one; omit and use the DOI if the paper has none; discard entirely if it resolves to a different paper.

Metadata mismatch: confirm the correct paper, not a similarly titled one; check preprint-vs-published discrepancies; use the most current metadata; note significant discrepancies (e.g., author order).

Claim not confirmed: if absent from the abstract, note "not confirmed from abstract — verify against full text"; never fabricate page numbers or section references.

Execution Mode: Dispatch vs. Inline

If the Agent/Task tool is available, dispatch the citation-verifier subagent (this plugin's agents/ directory for Claude Code, .claude/agents/ for Cowork), requiring tool-call evidence for every VERIFIED identifier and spot-checking at least 30% independently (see Multi-Agent guardrails below). If the tool is unavailable, run the workflow inline, unchanged.

Multi-Agent Citation Workflows

Subagents can confabulate verification reports: a claim of "all 12 references verified" with no tool-call evidence has verified nothing (references/history.md). Guardrails: (1) any subagent verifying citations must include its actual PubMed MCP calls (queries, get_article_metadata results) — a narrative-only report is not verification; (2) the parent spot-checks at least 30% of PMIDs directly, and re-verifies all on any failure; (3) no transitive trust — never accept a subagent's claim that it "searched PubMed" without independent verification; (4) before delivery, the delivering agent must hold direct tool-call evidence for every PMID in the document.

The dispatch-prompt template for verification subagents (restates guardrail 1) is in references/history.md.

Output Format (Audit Requested)

CITATION VERIFICATION REPORT

Verified: [X] / [Total]
Flagged:  [Y] (metadata issues or unconfirmed claims)
Failed:   [Z] (not found)

| # | Citation | Status | PMID Verified | Issue (if any) |
|---|----------|--------|---------------|----------------|
| 1 | Shah et al., PCCM 2024 | ✓ Verified | ✓ via MCP | — |
| 2 | Jones et al., CCM 2023 | ⚠ Flagged | ✗ Wrong paper | PMID resolves to unrelated article |

Integration with Other Skills

• fabrication-audit: owns non-academic verifiable specifics (scope split above); the two skills compose without overlap.
• clinical-citation-audit: the pre+post-flight gate wrapping this skill for the six clinical-writing skills (manuscript-reviewer, scientific-writing, literature-review, clinical-reports, clinical-evidence-debate, clinical-teaching-deck); that skill enforces, this skill supplies the methodology, and defers block formatting to its audit summary when invoked from the pipeline.
• manuscript-reviewer, review-response, literature-review, icu-clinical-consult, interesting-articles-formatting: each hands its citation-bearing content through this skill's workflow (icu-clinical-consult routes free-standing doses to fabrication-audit instead).
• writing-anti-ai: vague attributions it flags ("Experts believe...") should be replaced with verified specific citations.

What This Skill Does NOT Do

• It does not override the user's citation choices; it flags problems and the user decides.
• It does not verify references the user provides verbatim, though it flags obvious metadata errors if noticed.

Web fallback

WebFetch and WebSearch stay first. If the Parallel Search MCP is installed (claude mcp add --transport http parallel-search https://search.parallel.ai/mcp) and WebFetch returns a 4xx/5xx, a paywall or bot wall, a redirect it cannot follow, or a page missing the target content, retry once with mcp__parallel-search__web_fetch (same URL, a narrow objective); if WebSearch returns nothing usable, retry with mcp__parallel-search__web_search. Parallel returns query-matched excerpts, not the full page, so a value missing from the excerpt is unverified, not absent. Never put PHI, unpublished manuscript or grant text, or credentials in a Parallel query, and never use it to get past a login or CAPTCHA gate. Name the tool that verified each item.

Feedback loop

Found a missed edge case, a wrong-shaped output, or a rule that misfires? Open an issue on this plugin's repository with the input and the output you expected. Do not edit this skill mid-run.

Versions

See references/history.md for version history and the incidents behind each rule.