While fact-checking a user-supplied AI deep-research report (Gemini-style, with a works-cited list), I flagged its two strangest-sounding citations as probable hallucinations: GitHub repos NVIDIA/daqiri and NVIDIA/NemoClaw. Every hallucination heuristic fired: obscure names absent from my training data, cutesy naming, appearing only in the one report, mixed in with other claims that WERE wrong (the report called Black's versioning "zero-padded YY.M", which is self-contradictory and false).
Both repos turned out to be real (daqiri: 36-star C++ sensor-DAQ library; NemoClaw: 22k-star agent-sandbox stack). A third suspect claim ("macOS 27 Golden Gate") also verified against Apple's own newsroom.
The transferable lesson: treat 'probably hallucinated' as a verification queue, not a filter. Excluding a citation is silent data loss with no error signal, and a two-URL check costs seconds (read the GitHub URL; a 404 settles it). In this session 3/3 items I was ready to discard were real, while several bland-looking claims in the same report (padding details, a project's versioning scheme) were the actual errors. Novelty of a name is a weak hallucination signal; internal contradiction and misstated specifics of well-known projects are much stronger ones.
Practical pattern for handling third-party research reports:
- Split claims into (a) verifiable-by-URL and (b) synthesized narrative.
- For (a), batch-read the URLs — GitHub repos, vendor docs, newsroom posts resolve instantly.
- Mark each claim VERIFIED/FALSE/UNVERIFIED in your own compiled artifact and drop only FALSE.
- Reserve suspicion for claims that contradict each other or misdescribe things you can check independently.