verify-citations

name	verify-citations
description	Full bibliography audit for an academic paper. Verifies every citation is (1) structurally correct in the .bib file, (2) actually cited in the paper, and (3) a real publication — author names, year, journal, and title verified via web search. Flags hallucinated or inaccurate references.
allowed-tools	["Read","Grep","Glob","Write","Agent","WebSearch"]

Verify Citations

Full three-layer audit: structural integrity → cross-reference → hallucination check.

Input: <bib_file> — path to .bib file. Defaults to paper/references.bib.

---

Workflow

Step 1: Extract all citation keys used in the paper

Grep all .tex files under paper/ for citation commands:

\cite{, \citet{, \citep{, \citeauthor{, \citeyear{, \citealt{, \nocite{

Collect every unique citation key referenced in the paper.

Step 2: Parse the .bib file

Read <bib_file> (default paper/references.bib) and extract for each entry:

• Citation key
• Entry type (@article, @book, @incollection, @working paper, etc.)
• Required fields: author, title, year, journal (or booktitle / publisher)
• Optional fields: volume, number, pages, doi, url

Step 3: Structural audit (no web needed)

For each bib entry:

Check	Severity	Flag
Citation key in paper but not in .bib	CRITICAL	Missing entry
Entry in .bib but not cited anywhere	INFO	Unused entry
Required field missing (author/title/year/journal)	MAJOR	Incomplete entry
Year outside plausible range (< 1900 or > current year + 1)	MAJOR	Implausible year
Author field not in Lastname, Firstname or Firstname Lastname format	MINOR	Formatting
Malformed characters or encoding issues	MINOR	Encoding
Duplicate citation keys	CRITICAL	Duplicate

Step 4: Hallucination / accuracy audit (web verification)

This is the core new capability. For every citation key that IS cited in the paper, dispatch a web-search verification agent.

Launch a general-purpose agent with the following prompt template for each batch of 5–10 entries (batch to reduce agent count):

You are verifying whether academic citations are real and accurate.
For each entry below, search the web (Semantic Scholar, Google Scholar, SSRN, NBER, journal websites)
to verify:
  1. Does a paper with this title actually exist?
  2. Are the authors correct (right names, right order)?
  3. Is the year correct?
  4. Is the journal/venue correct?
  5. Is the DOI or URL resolvable (if present)?

For each entry, return one of:
  - VERIFIED: all fields confirmed
  - LIKELY OK: paper exists but minor discrepancy (e.g. slightly different title spelling)
  - SUSPECT: paper may not exist or fields are wrong — explain what you found
  - HALLUCINATED: paper definitively does not exist with these details

Entries to verify:
[paste bib entries here]

Return a structured table: Key | Status | Notes

Run verification agents in parallel (max 3 agents simultaneously, each handling a batch).

Step 5: Compile the report

Save to quality_reports/citation_audit.md using this format:

# Citation Audit Report
**Date:** YYYY-MM-DD
**Bib file:** paper/references.bib
**Paper:** paper/main.tex

## Summary
- Total entries in .bib: N
- Total citations used in paper: N
- Unused entries: N
- CRITICAL issues: N
- MAJOR issues: N
- Hallucination suspects: N

## CRITICAL Issues (must fix before submission)
[List with bib key, issue, fix]

## MAJOR Issues
[List with bib key, issue, recommendation]

## Hallucination / Accuracy Flags
| Key | Status | Filed As | What Web Search Found |
|-----|--------|----------|-----------------------|
...

## Unused Entries (informational)
[List of keys in .bib but not cited]

## VERIFIED entries
[Count only — full list omitted for brevity]

## MINOR Issues
[List]

Step 6: Present summary to user

Report:

• Total citations audited
• Any CRITICAL or hallucinated entries (full detail)
• Any MAJOR structural issues
• Count of verified entries
• Path to full report

---

Principles

• Web-first for hallucination. Do not rely on the model's training knowledge to verify a citation — always search the web. Training data can itself contain hallucinated references.
• SSRN and NBER first for working papers; journal websites for published articles; Semantic Scholar as a broad fallback.
• Flag, don't delete. The skill reports issues; the author decides what to fix.
• Conservative on HALLUCINATED. Only mark as HALLUCINATED if multiple searches find no evidence of the paper. Use SUSPECT for ambiguous cases.
• DOI is ground truth. If a DOI is present and resolves correctly, treat the entry as VERIFIED even if other fields have minor discrepancies.

---

Files to scan (default)

paper/**/*.tex
paper/appendix/*.tex

Bibliography location (default)

paper/references.bib

Output

quality_reports/citation_audit.md