ref-check — careful reference verification

The iron rule, and why it exists

NEVER guess or trust-from-memory an arXiv ID, journal reference, year, author list, page range, or title. Every factual field is verified against a fetched authoritative record before an entry is added, approved, or "fixed." Training-data recall of bibliographic details is a hypothesis to verify, never a source. If a field cannot be verified after trying at least two sources, it is FLAGGED, not silently kept or invented.

The rule is this severe because bibliographies are not copied from sources — they are copied from other bibliographies. An error, once printed, outlives its origin: the next hundred authors copy the entry, not the paper, and the error acquires the authority of repetition. A language model trained on that literature has learned the majority text, which for a propagated error IS the error — model recall of a reference reproduces the folklore version with full confidence. This is why "it looks right" is worthless here, and why an entry that half the field cites one way can still be wrong. Verification means fetching the record closest to the publisher and comparing field by field; nothing else counts.

A second consequence: an entry that a model drafted from memory is not a degraded citation — it is a fabrication with correct formatting. The observed pattern (see traps below) is a real identifier carrying invented co-authors, a paraphrased title, or another paper's journal block. Plausibility is the failure mode, not the reassurance.

Authoritative sources, by literature

The universal spine is CrossRef / the publisher's DOI landing page — canonical journal, volume, pages, year for anything with a DOI. Around it, use the registry native to the literature:

1. INSPIRE-HEP (physics/hep) — fetch BibTeX directly: https://inspirehep.net/api/literature?sort=mostrecent&size=1&q=arxiv%3A<ID>&format=bibtex (URL-encode : as %3A; old-style IDs URL-encode the slash: q=arxiv%3Ahep-ph%2F<YYMMNNN>). Cross-check the arXiv abs page.
2. arXiv abs page (https://arxiv.org/abs/<ID>) — title, full author list, journal-ref line. v1 titles sometimes differ from the published title; the published title wins when the entry cites the journal.
3. CrossRef / publisher DOI page — the arbiter for journal fields.
4. zbMATH / MathSciNet / Project Euclid / numdam — mathematics.
5. ADS — astronomy; PubMed/PMC — life sciences; DBLP — computer science (conference/proceedings metadata that CrossRef often garbles).
6. Zenodo / the software's own CITATION file — software and datasets.
7. Gallica / archive.org / WorldCat / publisher record — classical books and pre-DOI literature.

Search aggregators (Google Scholar, Semantic Scholar) are for FINDING a work, never for verifying it. They merge records, inherit upstream errors, and invent metadata for scraped PDFs. Once found, verify against the native registry or publisher.

Sleep ~2s between external calls; try a second source before declaring anything unverifiable. When two authoritative sources disagree, the one closest to the publisher wins for journal fields, and the disagreement itself goes in the audit log — it usually marks a folklore error in the losing source.

Field-by-field comparison (semantic, not textual)

• Authors — every author present, correct order, correct spelling including diacritics. TeX escapes vs unicode ({\'e} = é) are equal. A missing, extra, or misspelled author = FIX. Watch homonyms: same-surname collaborators are routinely merged into one person, and given names swapped between them.
• Title — word-level. Brace/capitalization/TeX-markup differences = formatting. Wrong, missing, or extra words = FIX.
• Journal, volume, number, pages, year — exact. Standard abbreviation vs full journal name = MINOR. Wrong volume/pages/year = FIX. Translated-journal pairs (Russ. Math. Surv./Uspekhi Mat. Nauk; Sov. Phys. JETP/ZhETF): either is fine if internally consistent; a mismatched year between the pair = FIX.
• arXiv ID, DOI — exact, and the DOI must resolve to the right work: fetch the landing page and compare its title. A syntactically valid DOI pointing at an erratum, a comment, or a different paper by the same group is a FIX that no string comparison catches.
• Entry type and roles — chapters in collections credit the chapter authors, with editors as editors; proceedings carry the conference year vs publication year distinction explicitly.

Classify each entry: OK (every factual field compared against a fetched record and matching) · MINOR (formatting only, no action) · FIX (any factual discrepancy — record field: ours -> authoritative + source URL) · UNVERIFIABLE (state exactly what was tried) · INTERNAL (self-references/companion notes — list, exempt).

Never mark OK without having actually fetched and compared. "Looks right" is not a verification.

Folklore traps (all observed in practice)

Each of these is a measured failure mode from real bibliography audits, not a hypothetical. They are the classes that survive casual checking because the entry looks healthy.

• Wrong years that circulate — a result becomes attached to the wrong year and the whole literature repeats it: e.g. a landmark irrationality theorem published in 2001 that half the citing literature dates to 2004 — the journal record, not the majority citation, is the evidence.
• Truncated author lists — the field remembers the famous pair and drops the rest (e.g. a four-author paper that the field routinely cites by its two most famous authors). Verify the FULL list even when the short form is universal.
• Merged authors — two distinct people fused into one: same surname, averaged initials, one entry where the record shows two names. The inverse also occurs — one author split into two by an initials variant.
• Phantom page numbers — a plausible page range attached to a paper that the journal published under an article number, or a first page copied from a different paper in the same issue. Article-number journals take the article number, not an invented range.
• Mislabeled keys — the entry under a key can be a different paper than the key name suggests (observed: a key naming one author trio holding a paper by a different trio). The key is part of the audit: check that key ≈ actual authors.
• Announcement vs final version — cite the cleanly citable version (book/journal), not a short announcement note, and say which in a comment.
• Preprint v1 title drift — published title wins.
• Prose initials vs keys — before renaming a key, grep every \cite usage AND read the surrounding prose: initials in prose may already refer to the correct authors even when the key is wrong.
• Another paper's journal block — an entry can carry the correct preprint ID/title with the journal/volume/pages/year of a different paper by the same authors (observed in practice). Verify the journal ref against the preprint server's own journal-ref line or CrossRef, not just "a" record by those authors.
• Descriptive phrase as title — entries written from memory often carry a paraphrase of the paper's subject instead of its actual title (observed five times in a single audit). Titles must be copied from the fetched record, never composed.
• Invented co-authors / wrong given names on real papers — LLM-drafted entries can attach plausible-but-wrong names to a correct identifier (observed: wrong given names on 3 entries, an extra co-author on 1, in one audit). Check every name against the record even when the ID resolves.
• Entirely wrong paper under a plausible key — key, note, and citing prose describe paper X while the entry's fields are paper Y (observed: an entry on one subject sitting under a key naming a different subject entirely). When key/note/prose disagree with the entry, read the citing prose to determine intent before fixing — the correct fix may be a new entry, not a field repair.
• A "verified" pass is not immune — one same-day verified addition carried wrong volume/pages from a hasty read of the abstract page (an off-by-one volume number). Volume/pages come from the journal-ref line or CrossRef, not from adjacent metadata on a page that lists several versions.

The claim check — does the cited work say what the sentence says?

A bibliography can be field-perfect and still lie, because the citation's real payload is the sentence attached to it. A full ref-check audits both layers. For every citation that carries factual weight — "X proved Y [12]", a number with a bracket after it, a definition attributed to a source, a "first shown in" — do this:

1. Open the work. Full text where available; the preprint version otherwise (note the version skew). The abstract alone verifies only abstract-level claims.
2. Locate the claim: the theorem number, equation, table, or page where the cited statement actually appears. Record the locator in the audit log — a claim check without a locator is an impression, not a check.
3. Verdict per claim: SUPPORTED (locator recorded) · DRIFTED (the claim is true but lives in a different work — often an earlier or later paper in the same series; fix the citation, not the prose) · INFLATED (the work proves a special case or a weaker form than the sentence asserts) · ABSENT (nothing in the work matches) · CONTRADICTED (the work says the opposite — it happens, usually via a sign, a convention, or a negation lost in paraphrase).
4. Report, never silently rewrite. For DRIFTED the fix is the citation; for INFLATED/ABSENT/CONTRADICTED the fix may be the prose, and that is the author's call. A checker who "fixes" meaning has exceeded the mandate.

The claim-layer traps mirror the field-layer ones: secondhand laundering (the citation and its misreading were both copied from an intermediary paper that itself never checked); review-article telephone (a survey's compressed paraphrase gets cited as if it were the original's claim); attribution creep ("first proved by" pointing at the famous paper when the record shows an earlier, obscurer proof); convention mismatch (the cited formula is right in the source's normalization and wrong in the citing paper's). None of these are visible from the bibliography file; they only fall to actually opening the work.

Process

Adding entries (small batches)

Verify each work per the rules above; match the house entry style of the target .bib; add under a dated comment block (%% ----- <purpose> (added <date>, verified)) with a per-entry % VERIFIED <date> <source-URL> comment; run a duplicate-key check (grep '^@' | extract keys | sort | uniq -d must be empty); never modify existing entries in the same pass. Also run a duplicate-work check: the same paper can already sit in the file under a different key, and two keys for one work will bite at citation time.

Full-file audit (report-then-fix pattern)

1. Chunk the .bib by entry boundaries (~15–20 entries per verifier; compute line ranges from grep -n '^@').
2. Verify in parallel, REPORT-ONLY — if your agent framework supports parallel subagents, fan the chunks out; verifiers never edit the .bib (concurrent writes corrupt it). Each returns structured per-entry verdicts with verbatim diffs and source URLs.
3. Fix serially — one pass (human-reviewed, or a single agent working from the approved diff list) applies FIX items only, keys unchanged unless a rename is explicitly approved.
4. Key renames are load-bearing — grep all \cite usages across the tex tree, update them atomically with the .bib, check prose initials (see traps).
5. Rebuild — full compile-plus-bibliography cycle (e.g. pdflatex+bibtex); zero errors, zero missing/undefined citations is the exit gate. A clean exit code is not the gate — read the log for warnings, and confirm the rendered bibliography actually carries the fixes (stale .bbl files survive successful-looking builds).
6. Record — stamp fixed entries % VERIFIED <date> <source>; log the audit (date, counts by verdict, fixes applied) wherever the project tracks provenance. The log is what lets the next audit skip re-fetching two hundred already-verified records — an unstamped verification evaporates.

Exit checklist

• Every entry has a verdict; zero entries skipped.
• Every FIX applied cites its source URL; every UNVERIFIABLE is flagged to the user, not silently retained.
• Every DOI resolved and its landing page compared, not just string-checked.
• Claim check done on every load-bearing citation, with locators (theorem/equation/page) recorded; DRIFTED/INFLATED/ABSENT/CONTRADICTED verdicts reported to the author, prose untouched.
• Duplicate-key and duplicate-work checks clean; all \cite keys in the tex tree resolve (bibliography log has no missing-entry warnings).
• Full rebuild passes with zero errors and zero undefined citations, and the rendered output carries the fixes.
• Audit logged with date, counts by verdict, and fixes applied.

Acknowledgment. This procedure leans entirely on open bibliographic infrastructure — INSPIRE-HEP, arXiv, Crossref, zbMATH Open, MathSciNet, NASA ADS, PubMed and DBLP; we are grateful to the people who maintain them. Work that relied on ADS should carry the acknowledgment its FAQ asks for ("This research has made use of the Astrophysics Data System, funded by NASA under Cooperative Agreement 80NSSC21M00561").