One command to detect hallucinated, broken, and retracted citations in your paper.
bibguard v0.2.0 paper.bib — 42 entries [ 1/42] vaswani2017attention ✅ crossref, dblp [ 2/42] he2016resnet ✅ crossref, s2 [ 3/42] fake_quantum_paper ❌ no match [ 4/42] wakefield1998mmr ⚠️ crossref ... [42/42] brown2020gpt3 ✅ s2, openalex ────────────────────────────────────────────────── ✅ 38 ⚠️ 2 ❌ 2 (42 entries in 67.3s) FAIL entries: ❌ fake_quantum_paper — phantom_doi ❌ halluc_gpt_ref_7 — verification ──────────────────────────────────────────────────
CLI + Python API. TeX cross-audit, duplicate detection, auto-fix .bib.
pip install bibguard
bibguard paper.bib
Zero dependencies. Works in Node.js 18+ and browsers. All APIs support CORS.
npx bibguard paper.bib
# or: npm install bibguard
Paste BibTeX into a popup. All verification runs client-side.
Download from GitHub
Paste BibTeX entries below. Verification runs entirely in your browser — no data leaves your machine.
arXiv, Crossref, DBLP, Semantic Scholar, OpenAlex. Falls back gracefully across sources.
DOIs and arXiv IDs that look valid but don't resolve — the strongest hallucination signal.
A phantom ID cannot be overridden by a similar search result. No false negatives on fabricated identifiers.
\cite{key} with no .bib entry, and orphan entries never cited. (Python version)
Corrected .bib with missing DOIs and eprint IDs filled in from API results.
TypeScript version is zero-dependency. All 5 APIs support CORS — no proxy needed.
58-case golden test set. Reproduce it yourself.
| Category | Metric | Result |
|---|---|---|
| Hallucinated (14 fabricated) | Detected as FAIL | 14/14 (100%) |
| Chimera (5 mixed-metadata) | Detected as ≥ WARN | 5/5 (100%) |
| Real papers (10) | False positive (FAIL) | 1/10 (10%) |
| Retracted (28) | Any issue flagged | 19/28 (68%) |
| Runtime | 58 entries | 95s (~1.6s/entry) |
| Python vs TypeScript | Agreement on 48 real refs | 90% (43/48) |
No API keys required. All support CORS for browser use.
Skill definitions for major AI coding assistants.
/bibguard paper.bib/bibguard paper.bibauto on .bib/.texbibguard --jsonbibguard handles L0 (existence & metadata verification). For semantic NLI, citation intent classification, graph anomaly detection, and Bayesian risk scoring, see IntegriRef — the full L0-L4 verification stack.