You help a medical researcher prevent reference hallucinations before submission.\nThis skill audits an existing manuscript or bibliography. It does not write\nto references/ or manuscript/_src/refs.bib. It does not discover new\nliterature; use /search-lit for discovery and /lit-sync for bib management.
.docx manuscripts inherited from\ncoauthors or external editors./sync-submission freezes a journal package..md, .docx, .bib, .txt, or .tsv.--offline: extract and classify references without API verification.--timeout N: HTTP timeout seconds.For markdown manuscripts using pandoc [@bibkey] citations, validate citation\nkeys first to catch undefined/unused keys before this audit. If you also use the\ncompanion manage-refs skill, run its check_citation_keys.py for this;\notherwise use your reference manager's citation-key check.
Then run verify_refs.py against the .bib to validate each entry against\nPubMed/CrossRef. The two checks are complementary: a citation-key check catches\nmis-keyed cites; verify_refs.py catches fabricated metadata.
Run the bundled script rather than verifying citations by memory:
\npython \"${CLAUDE_SKILL_DIR}/scripts/verify_refs.py\" manuscript/manuscript.md --project-root .\nFor hooks or quick manual runs, use the wrapper:
\n\"${CLAUDE_SKILL_DIR}/scripts/verify_cli.sh\" manuscript/manuscript.md --offline\nManual pre-submission strict run (Phase 1A.5):
\n\"${CLAUDE_SKILL_DIR}/scripts/verify_cli.sh\" manuscript/index.qmd --strict\n--strict forbids --offline and exits non-zero on any UNVERIFIED row.\nFull checkpoint protocol: references/manual_checkpoint_guide.md.
The script uses DOI, PMID, CrossRef, PubMed E-utilities, and OpenAlex where\navailable. If network verification fails, it records UNVERIFIED rather than\nsilently passing.
OpenAlex tertiary index (existence recovery). PubMed covers only biomedical\nliterature and CrossRef's conference-proceedings coverage is uneven, so\nNeurIPS / ICLR / ACL-style citations — common in medical-AI manuscripts — fall\nthrough both and would be marked UNVERIFIED. After the PubMed and CrossRef tiers,\nthe script consults OpenAlex (https://api.openalex.org, free, no API key) only\nwhen no authoritative author list was obtained yet (so a reference already\nresolved by PubMed/CrossRef incurs no extra call). It resolves by DOI when present,\notherwise by a title search guarded by a token-similarity threshold so a fabricated\ntitle cannot earn a spurious OK. This is the free analogue of the second index\n(e.g. Scopus) that journal submission portals run alongside CrossRef. OpenAlex\ndisplay names carry no structured family/given split and mix First Last with\nLast, First forms, so OpenAlex-sourced authors support an existence check plus a\ntolerant first-author membership check, but never drive the strict positional or\nauthor-count MISMATCH (those stay reserved for PubMed efetch / CrossRef). An\nOpenAlex miss is recorded as UNVERIFIED, never FABRICATED. Pass --no-openalex\nto restrict verification to PubMed + CrossRef.
| Artifact | \nPath | \nPurpose | \n
|---|---|---|
| Audit JSON | \nqc/reference_audit.json | \nMetadata audit output — row-level status (OK/MISMATCH/UNVERIFIED/FABRICATED), counts, cited_authors[]/actual_authors[], duplicate_findings[], submission-safe flag, full records | \n
v1.2.0 (2026-05) adds duplicate_findings[] to the audit JSON. Verbatim PMID or DOI duplicates within the reference list are flagged as MAJOR findings (resolves /peer-review Phase 2A P7). DOI normalization strips https://doi.org/, http://dx.doi.org/, doi: prefixes plus trailing slashes before comparison so https://doi.org/10.x/abc/ and 10.x/abc collapse to one key. Both submission_safe and fully_verified now require duplicate_findings to be empty.
v1.3.0 (2026-05) extends the author cross-check from first-author-only to the full author list and bumps schema_version to 4. For BibTeX inputs, every cited author family name is compared index-by-index against the authoritative source, and the cited-vs-source author counts are compared. PubMed efetch.fcgi (XML full record) is the truth source when a PMID is present — it is authoritative for given/family names where CrossRef is not (a documented case where CrossRef returned a wrong given name that PubMed efetch corrected). Records now carry cited_authors[], actual_authors[], cited_author_count, and actual_author_count. A correct first author does not establish that the remaining author names are authentic. Plain-text / TSV inputs, which cannot be parsed into a confident full list, degrade gracefully to the first-author check.
Removed in Phase 1A.2 (per docs/artifact_contract.md):
references/verified_references.tsv — record-level details now live inside reference_audit.json under records[].references/library.bib — never this skill's concern. /search-lit produces candidates; /lit-sync (via Better BibTeX) writes manuscript/_src/refs.bib.Sole-writer enforcement: scripts/validate_project_contract.py will flag any references/* file written by this skill as drift.
scripts/verify_refs.py.qc/reference_audit.json.FABRICATED and MISMATCH rows first (from records[]).duplicate_findings[] entries (verbatim PMID/DOI duplicates — cite renumbering required).UNVERIFIED rows remain, list them as manual checks and do not call the\nmanuscript fully submission-safe. Rows with note = \"pagination_placeholder\"\n(e000–e000 / in press / TBD / forthcoming) need the citation resolved\nbefore submission; /self-review Phase 2.5c decides whether any is a P0 blocker.records[] in chat — do not write a TSV.FABRICATED.UNVERIFIED references.MISMATCH. First-author mismatches get\nnote = \"first-author hallucination suspected\"; #2..#N family or count\nmismatches get note = \"non-first-author hallucination or count mismatch\".\nThis catches the LLM failure mode where a real DOI is paired with invented\nauthor names anywhere in the list, not just the lead author. Intentional CSL\net-al truncation (cited fewer than source) can be silenced per-entry with a\nBibTeX _audit_truncated = <N> field.duplicate_findings[]. submission_safe == true requires the list to be\nempty. Resolves /peer-review Phase 2A P7.e000–e000, in press, TBD, or forthcoming\nis not yet a fully citable record. Each is marked UNVERIFIED with\nnote = \"pagination_placeholder\" (a would-be VERIFIED record is downgraded; a\nworse status is left unchanged). verify-refs is manuscript-agnostic and does not\njudge centrality — it only flags. The escalation call (is this a method- or\nheadline-load-bearing citation, hence a P0 submission blocker?) is made by\n/self-review Phase 2.5c, which has the manuscript in hand.Classification note — citation-metadata confusion is not fabrication. Digits\nin a DOI suffix sometimes look like a journal article number but differ from the\nreal one (e.g., a DOI tail \"77196\" against article number 26068, or a \"60466-1\"\nsuffix against article 6274). This is cosmetic metadata confusion, not a\nfabricated reference: do not record such rows as FABRICATED when the DOI/PMID\nresolves and the authors match. A genuine FABRICATED verdict requires a\nnon-resolving identifier or an author cross-check failure (Gate 4), not a\nmismatch between a DOI suffix and an article number.
Two failure patterns motivate the author checks: a real DOI can be paired with\nthe wrong first author, and a correct first author can be followed by fabricated\nco-author names. DOI resolution and first-author agreement alone cannot verify\nthe full author list.
\nefetch.fcgi (XML) when a\nPMID is present, falling back to CrossRef (DOI) and then PubMed esummary.\nefetch is preferred because CrossRef is unreliable for given names.cited_authors[],\nbalanced-brace aware, LaTeX-accent tolerant) and compared family-by-family and\nby total count against actual_authors[]._audit_truncated = <N> to downgrade the count mismatch to a note.verify_refs.py answers whether a reference is real and whose it is. It cannot answer\nwhether the sentence citing it is true of it, and a citation can be perfectly real while\nthe claim attached to it is not. That gap is where the failure lives: the DOI resolves, the\nauthors match, the reference list renders, and the sentence is still wrong.
scripts/check_claim_fidelity.py checks the claims that have a checkable answer, against\nfull texts you have already downloaded and converted (/fulltext-retrieval produces exactly\nthat layout — it never fetches anything itself):
python3 \"${CLAUDE_SKILL_DIR}/scripts/check_claim_fidelity.py\" \\\n --manuscript manuscript/manuscript.md \\\n --fulltext-dir fulltext/ --bib manuscript/_src/refs.bib \\\n --out qc/claim_fidelity.json --strict\n| Verdict | \nSeverity | \nFires when | \n
|---|---|---|
CITED_QUOTE_ABSENT | \nmajor | \nQuoted text attributed to a source is not in it in any reading order. | \n
CITED_QUOTE_UNRESOLVED | \nprompt | \nThe quote matched only with foreign tokens wedged in, or a word or two missing — the signature of a dirty extraction, not of a fabrication. Look; do not assume. | \n
ATTRIBUTION_UNSUPPORTED | \nprompt | \nNot one content word of the attributed claim appears in the source, in any form. Paraphrase normally keeps at least one of the source's own terms. | \n
ORDINAL_CLAIM_UNSUPPORTED | \nprompt | \n\"reports three strategies [12]\" where the source discusses that noun but never that count near it. | \n
Only the quote verdict can fail --strict. Everything else is a prompt to go read the\nsource, because paraphrase is legitimate and a gate that blocks on it would be turned off.
Read the \"not checked\" lines. A citation with no full text on disk is reported as\nunresolved and never guessed at, and a source whose extracted text is an abstract is reported\nas too short to judge — absence proves nothing against an abstract. Silence from this\ndetector means \"nothing checkable was wrong\", which is not the same as \"everything is right\".
\nThe same qc/claim_fidelity.json now includes evidence_rows: recognized prose\nsentence/citation pairs, manuscript coordinates, source-text and PDF hashes,\nadvisory retrieval identity, and a separate assessor-entered comparison. Initial\nrows are not_assessed, even when bibliographic status is OK and no probe fires.
python3 \"${CLAUDE_SKILL_DIR}/scripts/check_claim_fidelity.py\" \\\n --manuscript manuscript/manuscript.md --bib manuscript/_src/refs.bib \\\n --fulltext-dir fulltext/ --retrieval-report pdfs/retrieval_report.json \\\n --reference-audit qc/reference_audit.json \\\n --out qc/claim_fidelity.json --evidence-table qc/claim_fidelity.md\nInspect the actual source before entering pages, excerpts, metric/unit/denominator,\npopulation, direction, and a named assessment. Neither equal numbers nor matching\nwords establish support. Record whether the assessor used AI assistance; do not\ndescribe an AI-generated assessment as human approval. Rerun with\n--reviewed-report qc/claim_fidelity.json to retain annotations. Changed inputs\nleave old assessments unresolved; unmatched rows remain in the JSON for review.\nThe Markdown table is a derived view, not a second editable evidence store.
See references/claim_evidence_workflow.md for field meanings, re-review steps,\nsource-identity limitations, and the difference between recorded and verified.
/fulltext-retrieval); claim fidelity reads converted text\noff disk so it stays deterministic and CI-runnable./search-lit or user approval./lit-sync after this audit.UNVERIFIED and keep it visible.