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.

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:

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.

Classify each entry: OK (every factual field compared against a fetched record and matching) · MINOR (formatting only, no action) · FIX (any factual discrepancy — record + 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.

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.

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:

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.

Verify each work per the rules above; match the house entry style of the target .bib; add under a dated comment block () with a per-entry comment; run a duplicate-key check ( 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.

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").