Skip to content

Bibliography audit — 2026-09-03, coordinator, RUN#

references.bib is the merge of the eight sweep bibliographies (508 entries parsed, 51 same-key duplicates dropped, 49 different-key/same-identifier entries folded into one canonical key each — the alias table is bib_aliases.md). 408 canonical entries.

The Scholar Sidekick audit service returned "not subscribed", so the same identifier-level check (real identifier paired with the wrong title — the Topaz et al. fabrication pattern) was run directly against the two registries:

keyed by entries resolved title agreement below threshold unresolved
arXiv id (eprint, or a 10.48550/arXiv. DOI) 278 278 (export.arxiv.org API, 5 batches) 1 — bottou-2013-counterfactual (cited with its subtitle; the API title is the short form; same paper) 0
DOI (Crossref works API) 92 91 6 — all subtitle truncations of books or a capitalised registry title: bollt-2001-misplaced, durrett-2019-probability, edelsbrunner-2009-computational-topology, pearl-2009-causality, robbins-1971-almost-supermartingales, villani-2009-optimal (same works) 1 — singh-2007-mapper, DOI 10.2312/SPBG/SPBG07/091-100 is not in Crossref (Eurographics digital library); carried [U], as the topology sweep already marked it
neither (books, 19th–20th-century papers, proceedings without DOI) 38 carried with the sweep's catalogue / ISBN / zbMATH / DBLP marks and its [V]/[U] tag

Title agreement is difflib ratio on lower-cased alphanumerics; thresholds 0.8 (arXiv) and 0.75 (Crossref). Every entry below threshold was read by eye and is the cited work under a shorter registry title.

What this audit does not establish: that any source supports the sentence it is cited for. That is the sweeps' [V] (abs page fetched, title matched) and [V-eq] (equation transcribed, hypotheses stated, instance run) marks, which are per-source in sweep_*.md; the paper carries the mark beside each citation that is load-bearing. The record's rule (99777ab:PRIOR_ART.md §1.2) stands: theorem numbers and equation numbers read through a rendered page should be re-checked against the compiled PDF before publication.