bib-verify
A fabricated reference is a research-integrity failure, not a typo. Typical LLM-era failure modes seen in
practice: invented co-author lists on a real paper; a placeholder like {Chen and others} + a journal name
propping up a row of a comparison table and a "nobody has done X" claim; a key that points at a different
paper by the same group; an ePrint number written from memory that belongs to another paper. When this repo's
own reading lists were drafted from memory, the verifier caught three such errors out of ~110 entries.
Procedure
- Run the checker (network needed; polite 1 req/s):
Resolution order: ePrint id (frompython3 skills/bib-verify/scripts/verify_bib.py refs.bib --json bib-verify.json > bib-verify-report.mdeprint=,url=,howpublished=...) → OAI record;doi=→ Crossref, then DataCite (Zenodo/LIPIcs/arXiv DOIs); otherwise a title search (DBLP if it answers, else Crossref). - Triage every non-
VERIFIEDrow by hand:status meaning action MISMATCHthe id/DOI resolves to a different title or author set fix the id or the metadata; if the cited claim was about the other paper, re-read and re-check the sentence that cites it NOT_FOUNDnothing plausible exists treat as fabricated: delete, or find the real source and re-cite; re-check every sentence that relied on it WEAKsimilar title only open the landing page, compare authors/year/venue manually UNCHECKEDoffline / service error re-run later; never ship with UNCHECKED rows - Lint flags (placeholder authors,
and othersonly, missing year,% UNVERIFIEDmarker, duplicate titles) must be cleared too. - For every changed entry, grep the paper for its key and re-read the citing sentence: a wrong reference often props up a wrong claim (a comparison row, a novelty statement, a parameter).
- Record the run in
LITERATURE.md(date, counts) and keepbib-verify-report.mdnext torefs.bib; thephase-gatescript checks that the report is fresh for the bib (the report'sbib-sha256:line must equal the hash of the currentrefs.bib; a--onlyrun writes no hash and therefore never vouches for the whole file) and clean.
Rules
- Prefer the published version (DOI) for the final paper; keep the ePrint id in
eprint=so it can be re-checked. - Never "fill in" missing authors from memory. Copy from the ePrint/DOI landing page.
- Anything added without checking gets a
% UNVERIFIEDline directly above the entry until verified. - Venue strings are not machine-verified; check them on the DOI landing page.
- Offline:
--offlinestill runs the static lint; the gate stays closed until an online run passes.
Files
scripts/verify_bib.py— dependency-free checker (--selftest,--selftest-online,--only k1,k2,--offline).