Nine citations on the wrong page, and one DOI that resolved to a stranger

A 59-entry reference catalog, drafted by models, two to five publications per entry, every one checked by a script that verifies title and authors against Crossref. All 99 pass.

Reading the 59 entries and asking one question per reference, “is this paper about this sensor”, found nine that were not. A study of interrupted work under code review. The truck-factor paper under dependency graphs. One mutation-testing paper under line coverage, another under diff coverage. A memory-safety blog post under static security analysis. Reproducible builds under incremental build correctness. None was invented. Every one exists, resolves, and is about the topic next door.

A citation checker verifies that a reference exists. It cannot verify that it belongs. Three of the nine were the right paper for a sensor the catalog does not have, which an earlier audit had already read as a tell: a paper cited and its sensor missing means the bibliography was assembled apart from the entries and nobody followed the link.

Then the repair did the same thing in the other direction. A DOI recalled for Licker and Rice 2019 resolved, on Crossref, to Motwani and Brun 2019. Same conference, sequence number 35 instead of 125, plausible to the last digit. Title search found the real one.

When agents write reference material the failure is not fabricated sources. The script catches those. It is real sources on the wrong page, and only a reader holding the page and the paper together catches that. Budget the read, and treat a recalled identifier as a claim to resolve, not a fact to type.