citation-verifier

Citation Verifier

Detect and verify citations in scientific documents to identify hallucinated, broken, or invalid references.

Purpose

AI-generated content sometimes includes plausible-looking but fake citations. This skill systematically extracts all citation identifiers from a document and verifies each one against authoritative sources, producing a detailed report with verification status and suggestions for fixing invalid citations.

When to Use

This skill should be invoked when:

• User asks to "verify citations" or "check references" in a document
• User suspects hallucinated citations in AI-generated content
• User wants to validate DOIs, URLs, or other identifiers in a paper
• User asks to audit a document for broken links or fake references
• User mentions "citation verification", "reference checking", or "DOI validation"

Supported Document Formats

1. Markdown (.md): Inline links [text](url), reference links [text][ref], bare URLs, DOIs
2. LaTeX/BibTeX (.tex, .bib): \cite{}, @article{}, DOI fields, URL fields
3. Org-mode (.org): [[url][text]] links, #+BIBLIOGRAPHY, cite links
4. Plain text (.txt): Bare URLs, DOIs, arXiv IDs, author-year patterns

Citation Identifiers Detected

DOIs (Digital Object Identifiers)

• Pattern: 10.\d{4,}/[^\s]+ or doi.org/10.\d{4,}/[^\s]+
• Example: 10.1038/nature12373, https://doi.org/10.1126/science.abc1234
• Verification: CrossRef API at https://api.crossref.org/works/{doi}

URLs to Papers

• Patterns: Links to known publishers and repositories
• Domains: nature.com, science.org, sciencedirect.com, springer.com, wiley.com, acs.org, rsc.org, pnas.org, cell.com, plos.org, mdpi.com, frontiersin.org, academic.oup.com, tandfonline.com
• Verification: HTTP HEAD/GET request, check for 200 status and paper metadata

arXiv IDs

• Pattern: arXiv:\d{4}\.\d{4,5}(v\d+)? or arxiv.org/abs/\d{4}\.\d{4,5}
• Example: arXiv:2301.07041, https://arxiv.org/abs/2301.07041v2
• Verification: arXiv API or direct URL check

PubMed IDs (PMIDs)

• Pattern: PMID:\s*\d+ or pubmed.ncbi.nlm.nih.gov/\d+
• Example: PMID: 12345678
• Verification: PubMed URL https://pubmed.ncbi.nlm.nih.gov/{pmid}/

ISBNs

• Pattern: ISBN[:\s]*[\d-]{10,17} (ISBN-10 or ISBN-13)
• Example: ISBN: 978-0-13-468599-1
• Verification: Open Library API https://openlibrary.org/isbn/{isbn}.json

Author-Year Citations

• Pattern: ([A-Z][a-z]+(?:\s+(?:et\s+al\.?|and|&)\s+[A-Z][a-z]+)?,?\s*\d{4})
• Example: (Smith et al., 2023), (Johnson and Lee, 2022)
• Verification: WebSearch to find matching paper (lower confidence)

Verification Procedure

Step 1: Read and Parse Document

Use the Read tool to load the document. Extract all citation identifiers using pattern matching:

DOI patterns:
- https?://(?:dx\.)?doi\.org/(10\.\d{4,}/[^\s\])"'>]+)
- doi:\s*(10\.\d{4,}/[^\s\])"'>]+)
- (10\.\d{4,9}/[-._;()/:A-Z0-9]+)  (bare DOI)

arXiv patterns:
- arXiv:(\d{4}\.\d{4,5}(?:v\d+)?)
- arxiv\.org/abs/(\d{4}\.\d{4,5}(?:v\d+)?)

PubMed patterns:
- PMID:\s*(\d+)
- pubmed\.ncbi\.nlm\.nih\.gov/(\d+)

URL patterns:
- https?://[^\s\])"'<>]+  (filter for academic domains)

ISBN patterns:

Step 2: Deduplicate and Categorize

Create a list of unique identifiers, categorized by type:

• DOIs
• arXiv IDs
• PubMed IDs
• ISBNs
• URLs (academic)
• Author-year citations (text-based)

Step 3: Verify Each Identifier

For each identifier, perform verification in order of reliability:

DOI Verification

1. Construct CrossRef API URL: https://api.crossref.org/works/{doi}
2. Use WebFetch to check the API
3. If successful, extract: title, authors, journal, year
4. If 404 or error: mark as INVALID

arXiv Verification

1. Construct URL: https://arxiv.org/abs/{arxiv_id}
2. Use WebFetch to verify page exists
3. Extract: title, authors, abstract snippet
4. If 404: mark as INVALID

PubMed Verification

1. Construct URL: https://pubmed.ncbi.nlm.nih.gov/{pmid}/
2. Use WebFetch to verify
3. Extract: title, authors, journal
4. If 404: mark as INVALID

ISBN Verification

1. Construct URL: https://openlibrary.org/isbn/{isbn}.json
2. Use WebFetch to check
3. Extract: title, authors, publisher
4. If 404: mark as INVALID

URL Verification

1. Use WebFetch to access the URL
2. Check for HTTP 200 and academic content indicators
3. Look for: paper title, authors, DOI on page
4. If unreachable or non-academic: mark as SUSPICIOUS

Author-Year Verification (lowest confidence)

1. Use WebSearch with query: "{author}" "{year}" paper
2. Look for matching papers in results
3. If found: mark as LIKELY VALID with source
4. If not found: mark as UNVERIFIED

Step 4: Generate Report

Produce a structured verification report:

Verification Status Definitions

• VALID: Identifier resolves to a real paper with matching metadata
• INVALID: Identifier does not exist or returns 404 (likely hallucinated)
• SUSPICIOUS: Could not fully verify; may be rate-limited, paywalled, or temporarily unavailable
• UNVERIFIED: Text-based citation that couldn't be confirmed (conservative approach)

Best Practices

1. Batch similar requests: Group DOI checks together to minimize API calls
2. Respect rate limits: Add delays between requests if hitting rate limits
3. Cross-reference: If a URL contains a DOI, verify the DOI directly
4. Context matters: Note where citations appear (methods vs. claims)
5. Report uncertainty: Always distinguish between "confirmed invalid" and "could not verify"

Output Suggestions for Invalid Citations

For each invalid citation, provide actionable suggestions:

• Wrong DOI format: "DOI appears malformed. Check for typos or extra characters."
• Non-existent DOI: "No paper found. This may be hallucinated. Search for the actual paper title."
• Dead URL: "URL returns 404. Try searching for the paper title on Google Scholar."
• Suspicious journal: "Publisher not recognized. Verify this is a legitimate source."
• Author-year not found: "Could not verify. Add DOI or URL for confirmation."

Example Verification Session

User request: "Verify the citations in my-paper.md"

Expected behavior:

1. Read my-paper.md
2. Extract all DOIs, URLs, arXiv IDs, etc.
3. Report: "Found 15 citations: 8 DOIs, 5 URLs, 2 arXiv IDs"
4. Verify each identifier using appropriate API/fetch
5. Generate report showing:
   • 10 valid citations with metadata
   • 3 invalid citations (404 errors) marked as likely hallucinated
   • 2 suspicious citations (timeouts) requiring manual check
6. Provide suggestions for fixing invalid citations

Limitations

• Rate limits: CrossRef and other APIs may rate-limit requests
• Paywalled content: Cannot verify full content behind paywalls
• New papers: Very recent papers may not be indexed yet
• Author-year citations: Low confidence without additional identifiers
• Non-English sources: Limited support for non-English citation formats
• Private/institutional URLs: Cannot access authenticated content

Related Skills

• literature-review: For conducting systematic literature searches
• scientific-reviewer: For reviewing scientific document quality
• scientific-writing: For writing with proper citations