alterlab-citation-verifier

name	alterlab-citation-verifier
description	Verifies that every entry in a bibliography ACTUALLY EXISTS by cross-checking it against four keyless public scholarly APIs (Crossref, OpenAlex, Semantic Scholar, arXiv) with a polite mailto identifier, resolving DOI/arXiv IDs, fuzzy-matching title and authors (difflib SequenceMatcher ratio >=0.70), flagging retractions marked in Crossref (update-to) or OpenAlex (is_retracted), and emitting per-entry JSON verdicts mapped to the AlterLab citation-hallucination taxonomy (TF/PAC/IH/PH/SH). Accepts BibTeX, a DOI/arXiv ID list, or free-form references; degrades gracefully offline by emitting 'unverified' verdicts and never silently passing. Use when the request mentions verify citations, check references, fabricated or hallucinated references, fake DOI, retraction check, bibliography audit, or reference existence check. Does NOT write or draft papers — for authoring a manuscript (whose citation-check mode inserts citations) prefer alterlab-paper-writer instead. Part of the AlterLab Academic Skills suite.
license	MIT
allowed-tools	Read Write Edit Bash(python:*) Bash WebSearch WebFetch
compatibility	No API key required — cross-checks four keyless public scholarly APIs (Crossref, OpenAlex, Semantic Scholar, arXiv) via WebFetch and `uv run python`; degrades gracefully offline to 'unverified' verdicts
metadata	{"skill-author":"AlterLab","version":"1.0.0","last_updated":"2026-06-06","depends_on":"alterlab-research-pipeline (shares the integrity taxonomy), alterlab-deep-research"}

Citation Verifier — Existence-Verify a Bibliography Against Public Scholarly APIs

The headline existence-verification skill: given a bibliography in any common form, it proves entry-by-entry whether each reference actually exists by querying four keyless public scholarly APIs, then maps each result to the canonical AlterLab citation-hallucination taxonomy. It is the deterministic, network-grounded companion to the LLM-driven integrity_verification_agent — where that agent uses WebSearch + judgment, this skill uses authoritative API records and a reproducible Python script, so the same input always yields the same verdicts.

Quick Start

Verify the citations in references.bib
Check whether these DOIs resolve to the papers I cited
Audit my bibliography for fabricated / hallucinated references
Does this reference list contain any fake citations or retractions?

→ Run scripts/verify_citations.py over the bibliography, read the JSON, then present a verdict table grouped by severity. Always state the offline/degraded status explicitly if the network was unavailable.

---

WHAT This Does

For each bibliography entry the script:

1. Parses the input (auto-detects BibTeX / DOI-list / free-form), extracting title, authors, year, venue, DOI, and arXiv ID.
2. Resolves identifiers — looks up the cited DOI/arXiv ID directly when present.
3. Searches by title as a fallback across all four sources.
4. Fuzzy-matches the cited title (difflib SequenceMatcher ratio, default threshold 0.70) and computes author-surname overlap.
5. Flags retractions marked in Crossref (update-to: retraction) or OpenAlex (is_retracted).
6. Emits a verdict per entry mapped to the taxonomy below, plus a repo-level summary.verdict (PASS / PASS_WITH_CONDITIONS / FAIL / UNVERIFIED).

The four sources (all keyless, polite-pool)

Source	Endpoint	Used for
Crossref	api.crossref.org/works	DOI resolution, metadata, retraction flag
OpenAlex	api.openalex.org/works	DOI + title search, is_retracted