citeguard

Retracted papers get cited for years after retraction. Andrew Wakefield's fraudulent 1998 paper linking the MMR vaccine to autism — retracted in 2010 — has been cited well over a thousand times since its retraction, by researchers who had no easy way to know. A 2026 JMIR study found that freely available AI tools "cannot reliably flag retracted literature," right as AI-assisted research and writing has exploded. citeguard closes that gap: check a citation against Crossref — free, no API key — before it goes in a paper, a summary, or a review.

Three surfaces, one verified detection algorithm:

Surface	For	Location
Python library + CLI	academic writers, scripts	src/citeguard/
MCP server	AI agents doing research/writing	mcp-server/
GitHub Action	CI on a lab's or journal's repo	action.yml

Why this is built the way it is

The detection logic isn't guessed at from documentation — it was built by querying Crossref's real API for known cases and reading the actual response shapes, then writing tests against the saved real responses (committed in tests/fixtures/, shared by both the Python and TypeScript implementations). Two real papers anchor the ground truth:

• Wakefield et al., 1998, The Lancet (the MMR-autism paper): the publisher's own Crossref metadata has no structured retraction data — an update-to check alone would silently miss it. What actually carries the signal is a separate updated-by field, where Crossref has backfilled the retraction (and an earlier 2004 correction) from the Retraction Watch database itself. Missing this field would have meant missing the single most famous retracted paper in medicine.
• Mehra et al., 2020, The Lancet (the Surgisphere-linked COVID/hydroxychloroquine paper): here the publisher did attach structured data directly, via update-to. A different field, a different provenance, same underlying fact.
• Watson & Crick, 1953, Nature serves as the clean control in every test suite — a definitely-real, definitely-not-retracted paper that must never be flagged.

So the checker looks at three independent signals, in order: update-to (publisher-asserted), updated-by (often Retraction-Watch-sourced, catching what publishers miss), and a title-prefix fallback ("RETRACTED:", "WITHDRAWN:", etc.) for older or unlinked cases with no structured metadata on either field. Each signal is mapped to the right severity — an "Expression of Concern:" title is not the same thing as a retraction, and earlier versions of this code collapsing that distinction was itself a bug caught by testing against real titles, not just synthetic ones. See src/citeguard/analyze.py for the fully-commented implementation.

Quick start

CLI:

pip install -e .
citeguard doi 10.1016/S0140-6736(97)11096-0
citeguard file references.bib --fail-on retracted   # for CI, see below

MCP server (add to your client's config, e.g. .mcp.json):

{ "mcpServers": { "citeguard": { "command": "node", "args": ["/path/to/citeguard/mcp-server/dist/index.js"] } } }

GitHub Action (in another repo's workflow):

- uses: wedo911/citeguard@main
  with:
    path: references.bib
    fail-on: concern   # never | retracted | concern | corrected

What this is not

• Not proof a paper's content is correct. It only checks retraction status, not whether a non-retracted paper's findings hold up.
• Not exhaustive. The title-prefix heuristic only catches the publisher conventions it's been tested against; a clean result means "no known signal found," not "guaranteed never retracted."
• Not a bulk-scraping tool. It's built for the size of a real bibliography (tens of citations), with a small fixed delay between Crossref requests and an optional persistent cache (src/citeguard/cache.py) — good API citizenship for a free public service, not a tool for scanning millions of DOIs.

Running the tests

# Python (55 tests, including against the real fixtures above)
pip install -e ".[dev]" && pytest -v

# MCP server (17 tests against the same real fixtures)
cd mcp-server && npm install && npm run build && npm test

Both suites are network-free and deterministic — they run against the committed real API responses, not live calls, so they're fast and don't depend on Crossref being reachable. Live end-to-end behavior (the actual CLI, the actual MCP tool, hitting the real API) was separately verified by hand during development; the GitHub Action additionally has its own CI job (action-smoke-test) that runs the real composite action against a known-retracted and a known-clean bibliography on every push, so the Action itself — not just the underlying library — is continuously verified against the live API.

Contributing

New signal types, additional publisher title conventions, and false-positive reports are all welcome. If you add a case, prefer adding it as a real, cited Crossref fixture over a synthetic one where possible — that's what caught the two real bugs this project's own test suite found during development (a URL-encoding bug in the DOI request path, and a BibTeX parser that could swallow an adjacent entry when parsing a malformed @comment block).

License

MIT — see LICENSE.