ghostcite

Catch ghost citations — right DOI, wrong author.

ghostcite is a deterministic, no-LLM command-line tool that cross-checks a bibliography's claimed author and year against CrossRef's canonical record for each DOI. It catches the dominant ghost-citation failure mode — a reference whose cited authorship doesn't match the paper the DOI actually points to — and flags retracted or expression-of-concern works along the way.

The problem

LLM-assisted writing (and plain copy-paste drift) routinely produces references that look right but attribute the cited DOI to the wrong authors or year. A manuscript cites "Li et al. 2024," but DOI 10.3390/plants13060869 is actually Chen et al. A reviewer catches it; an automated check catches it first.

Does the metadata you wrote for this citation match what CrossRef says the DOI actually is?

No model, no API key, no GPU, no download — just CrossRef's REST API and a comparison. It runs in seconds on any CI runner, where the heavyweight LLM-based checkers can't (those need a local GPU — see Related work).

Install

pip install ghostcite          # into the current environment
pipx install ghostcite         # isolated CLI install (recommended)
uv tool install ghostcite      # if you use uv

Usage

ghostcite refs.bib                                    # check a BibTeX file (or .md / DOI list)
ghostcite refs.bib --cross-check pubmed               # corroborate against PubMed
ghostcite refs.bib --cross-check pubmed,openalex      # two independent sources of truth
ghostcite refs.bib --badge badge.json                 # write a shields.io citation-health badge
ghostcite refs.bib --json                             # machine-readable output (for CI)
ghostcite refs.bib --fail-on author,year,retraction   # tune the CI gate
cat refs.bib | ghostcite -                            # read from stdin

Input format is auto-detected (BibTeX, Markdown reference list, or bare DOI list); override with --format {auto,bibtex,markdown,doi}.

Real example — refs.bib cites "Li (2024)" for a DOI CrossRef says is Chen:

$ ghostcite refs.bib
ghostcite: 1 entries, 1 with DOIs
  ✗ A  L1  Li (2024)  →  DOI resolves to Chen (2024) — possibly wrong DOI  [10.3390/plants13060869]
  1 A
$ echo $?
1

All flags & the anatomy of a finding

  ✗ A   L1    Li (2024)        →  DOI resolves to Chen (2024)…   [10.3390/plants13060869]
  │ │   │     │                    │                               │
  │ │   │     │                    │                               └─ DOI that was checked
  │ │   │     │                    └─ what CrossRef actually records
  │ │   │     └─ what you cited (claimed first author + year)
  │ │   └─ source line in your bibliography
  │ └─ tier: A author · T title · B year · C cosmetic · R retraction · U unresolvable · V venue
  └─ glyph: ✗ fails CI · ⚠ retraction · · informational

• --cross-check pubmed — adds PubMed/NCBI as a second source of truth. When PubMed backs CrossRef a finding is annotated ↳ corroborated by PubMed; when PubMed instead agrees with what you cited, it's flagged as a CrossRef↔PubMed conflict (the tier is kept so you don't silently trust either source). PubMed can also raise a finding CrossRef missed, or supply a record for a DOI absent from CrossRef. Optional --ncbi-email / --ncbi-api-key (or NCBI_EMAIL / NCBI_API_KEY) follow NCBI E-utilities etiquette and unlock a higher rate limit; neither is required. --cross-check accepts a comma-separated list: --cross-check pubmed,openalex runs both PubMed and OpenAlex in the same pass (OR-combining retraction, mirroring the PubMed corroborate/conflict/raise logic). Optional --openalex-mailto / OPENALEX_MAILTO for the OpenAlex polite pool.
• --badge <path> — write a shields.io endpoint badge JSON to <path>. Green when the run is clean; red with a count when findings at the current --fail-on threshold are present. Warn-only tiers (C, U, V) do not turn the badge red. Serve the file from any static host and embed it with: https://img.shields.io/endpoint?url=<raw-url-to-badge.json>.
• --no-doi-probe — skip the best-effort HEAD https://doi.org/<doi> probe that ghostcite performs when a DOI is absent from CrossRef. The probe distinguishes "dead/fabricated DOI" from "resolves but not in CrossRef" and refines the Tier U message; pass --no-doi-probe for fully offline or air-gapped runs.
• --max-rps <n> — cap outbound requests per second. ghostcite already self-throttles to CrossRef's advertised rate limit (read from the response headers); --max-rps lets you be more conservative (the stricter of the two wins).
• --color {auto,always,never} — colorize the tier glyphs. auto (default) colorizes only on a TTY. NO_COLOR is honored and wins even over always. --json output is never colorized.
• stdin (-) — pass - as the filename to read from stdin, e.g. cat refs.bib | ghostcite - or ghostcite - --format doi < dois.txt.
• --dry-run — parse + classify + count only, no network.

See examples/ for ready-to-run sample inputs and captured output.

How it works

flowchart TD
    A["Citation: claimed author + year (+ DOI)"] --> B{"Has DOI?"}
    B -- yes --> C["GET CrossRef /works/{DOI}"]
    B -- no --> D["CrossRef bibliographic search<br/>(low-confidence)"]
    C --> E{"DOI resolves?"}
    E -- no --> U["Tier U — unresolvable"]
    E -- yes --> F["Compare claimed vs. canonical record"]
    D --> F
    F --> G{"First-author surname matches?"}
    G -- no --> TA["Tier A — author mismatch"]
    G -- yes --> H{"Year matches?"}
    H -- no --> TB["Tier B — year mismatch"]
    H -- yes --> OK["OK"]
    C --> R{"Retracted / expression of concern?"}
    R -- yes --> TR["Tier R — retraction (orthogonal)"]
    F -. "--cross-check pubmed" .-> P["PubMed second opinion"]

Loading

No language model is involved at any step. ghostcite resolves each DOI at CrossRef (and optionally PubMed), then does a pure, deterministic comparison of the claimed first-author surname (Unicode-folded, punctuation-stripped) and year against the canonical record, plus a retraction / expression-of-concern check. Only the HTTP client touches the network, via CrossRef's polite pool (a descriptive User-Agent with the project URL, never a personal email).

Severity tiers, input formats & exit codes

Tier	Meaning	Fails CI?
A	author-mismatch — claimed first author isn't in CrossRef's authors	Yes
T	title-mismatch — DOI resolves to a different paper (identifier hijack)	Yes
B	year-mismatch — author matches, claimed year differs	Yes
C	cosmetic — matches only after diacritic/initials fold (Bürger≈Burger)	No (info)
R	retraction / expression-of-concern per CrossRef	Yes (fires regardless of A/B/C)
U	unresolvable — DOI 404s, or no-DOI entry search was inconclusive	No (warn)
V	venue-mismatch — cited journal/venue differs from CrossRef's record (opt-in, abbreviation-tolerant)	No (info, opt-in)
OK	first author + year match	—

Tier T catches "identifier hijacking": the DOI resolves, but to a different paper than the one cited — even when the claimed author coincidentally matches the real first author. It compares the claimed title against CrossRef's canonical title with a conservative token overlap, so subtitle/formatting variance on a correct citation does not trip it. (When the author also mismatches, the wrong-DOI signal is folded into the Tier A finding instead, to avoid double-reporting.)

Format	Detection	Yields claimed author/year?
BibTeX	@article{…} / @…{…} entries	Yes (author, year, doi, title)
Markdown	bullet refs - **AuthorList (YYYY).** … 10.x …	Yes
DOI list	newline-delimited bare DOIs / doi: / DOI URLs	No — lookup + retraction sweep only

Exit code	Meaning
0	clean — no findings at or above the fail threshold
1	findings present at/above the threshold
2	tool error (network down, unparseable input, …)

--fail-on (default author,title,year,retraction) selects which tiers force exit 1; --fail-on none runs as a passive reporter. Tiers C and U never force exit 1.

Use it in CI

A clean run is quiet and exits 0:

Drop in the composite GitHub Action:

- uses: musharna/ghostcite@v1
  with:
    paths: paper/refs.bib
    fail-on: "author,title,year,retraction"

…or the pre-commit hook:

repos:
  - repo: https://github.com/musharna/ghostcite
    rev: v0.4.0
    hooks:
      - id: ghostcite
        # Staged .bib/.md files are appended automatically; args carries flags only.
        args: [--fail-on, "author,title,year,retraction"]

Either way, a finding at or above the --fail-on threshold returns a non-zero exit, blocking the merge or commit before submission.

Offline / reproducible retractions

By default ghostcite reads retraction status from CrossRef live. For a deterministic, byte-reproducible retraction check — with broader coverage than CrossRef's own flags — point it at a Retraction Watch snapshot:

ghostcite fetch-retractions --mailto you@org      # → ~/.cache/ghostcite/retractions.csv
ghostcite refs.bib                                # auto-uses the cached snapshot
ghostcite refs.bib --retraction-db vendor/rw.csv  # or pin an exact snapshot in CI
ghostcite refs.bib --retraction-db none           # force live CrossRef

When a snapshot is active it is the authoritative retraction source (the live CrossRef signal is skipped) and the report names the snapshot in use. The byline/year check still queries CrossRef live — this makes the retraction tier offline and reproducible, not the whole run. Pin a committed snapshot in CI for byte-reproducible results.

Retraction data: Crossref + Retraction Watch (The Center for Scientific Integrity). ghostcite downloads it on demand and does not redistribute it; see NOTICE.

Opt-in semantic claim-support (BYO backend)

ghostcite's defaults are deterministic and need no GPU. If you also want "does the cited paper actually support this sentence?", bring your own LLM:

ghostcite --claims claims.json \
  --semantic-backend openai \
  --semantic-base-url http://localhost:11434/v1 \
  --semantic-model llama3.1

claims.json is a list of {"claim": "...", "doi": "10.x/y"}. Results are abstract-grounded and marked non-deterministic; they never fail CI unless you add --fail-on support.

Scope & limitations

ghostcite's deterministic core checks metadata correctness (does the DOI's record match what you wrote), not deep claim support. It does offer an opt-in, abstract-grounded claim-support check (--semantic, bring-your-own LLM backend, still no local GPU — see above), but that stays intentionally lightweight: full-text entailment, internal-consistency, statistics, and figure checks are a heavier, GPU-bound concern best served by sciwrite-lint. ghostcite deliberately stays on the deterministic, no-GPU side so it can gate every commit and CI run in seconds. It does no auto-fixing and no citation-style linting. CrossRef is the source of truth; --cross-check pubmed adds PubMed as an optional second opinion, and --cross-check pubmed,openalex layers in OpenAlex as well.

• CrossRef stores particle surnames inconsistently (van der Berg vs Berg), so a correctly-cited prefixed surname can rarely produce a Tier A false positive.
• No-DOI entries are resolved by best-effort bibliographic search and flagged low-confidence — treat those as hints, not verdicts.
• Some preprints, datasets, and protocols carry no author metadata in CrossRef and surface as Tier U rather than a mismatch.

Related work & FAQ

ghostcite's niche is deterministic, no-GPU, CLI-first checking focused on the byline-mismatch failure mode (right DOI, wrong author/year) plus retraction flagging — built to run unattended, in seconds, on any CI runner. Its core never calls an LLM (an opt-in --semantic layer adds abstract-grounded claim support via a bring-your-own backend). The heavyweight LLM-based linters below verify more (full-text claim support, internal consistency) but need a local GPU and minutes-to-tens-of-minutes per paper; ghostcite is the lightweight gate you can run on every commit.

Tool	What it does	How ghostcite differs
sciwrite-lint (arXiv 2604.08501)	Local-LLM scientific-writing linter — byline + retraction plus full-text claim-support entailment, internal-consistency, stats, and figure checks	a capability superset, but needs an NVIDIA GPU (16 GB+) and runs minutes-to-30 min/paper; ghostcite is no-GPU, deterministic, sub-second, and CI-native — its --semantic check is abstract-only and backend-optional. Complementary: ghostcite gates every commit, sciwrite-lint deep-audits before submission
RefChecker	LLM-powered web-search reference validator	ghostcite is no-LLM, deterministic, and CI-safe (no model, no API key)
claude-skill-citation-checker	A Claude Code skill for an LLM agent	ghostcite is a standalone CLI + Action — no agent or LLM host needed
BibTeX Verifier	In-browser BibTeX checker	ghostcite is scriptable from the CLI and also flags retractions
CERCA	Java / AGPL citation checker	ghostcite is Python / MIT / pip install-able
scite Reference Check	Commercial, PDF-oriented, retraction focus	ghostcite is free / open-source, BibTeX-native, and catches byline mismatch
doimgr	Formats and manages DOIs (doesn't validate)	ghostcite verifies byline and retraction status, not just formatting

Does it call an LLM? Not by default. The core byline / year / retraction checks are a deterministic comparison of the metadata you wrote against CrossRef's (and optionally PubMed's) canonical record — no model, no prompt, no API key. Only the opt-in --semantic claim-support layer calls an LLM, and only through a bring-your-own backend you configure.

Will it hit rate limits? It self-throttles to CrossRef's advertised rate limit (read from the live response headers); use --max-rps to be more conservative.

Does it catch fabricated DOIs? Indirectly — a DOI absent from CrossRef surfaces as Tier U. ghostcite performs a best-effort HEAD https://doi.org/<doi> probe to distinguish "dead/fabricated DOI" (does not resolve at all) from "resolves but not indexed in CrossRef" and surfaces that distinction in the Tier U message. Pass --no-doi-probe to skip the probe for fully offline or air-gapped runs. The core check is byline-vs-DOI consistency, so it catches the common case of a real DOI attached to the wrong citation.

License

MIT — see LICENSE.