Aadinarayana “Aadi” Varma, Nexus Score
Community highlights, Session1, crossref2026
Aadi Varma hid and restored author, institution, funder, reference and text-access links in OpenAlex records. AI systems could only credit work when the matching link was visible or recoverable, and restoring a different kind of link gave 0 correct answers in 469 tests. Without the right link, the systems invented answers, refused or ran out of tool budget, and web search didn’t recover hidden author links.
Thank you for sharing this, Aadi, and Rosa for posting it. The finding that one kind of link cannot stand in for another is a useful one to show journal editors.
Two questions from the side of small journals that deposit their own metadata:
In the author test the links were removed from the OpenAlex records. Did you look at the common real-world case where an ORCID iD is printed on the article page or in the PDF but never deposited? Your note that web search did not recover the hidden author links suggests the printed iD does not help, but I may have missed it in the preprint.
Did restoring the reference list help the systems credit the cited works, or only the citing article?
Thank you — we will point editors to the preprint.
We didn’t test printed-but-undeposited ORCIDs directly. However, none of our 400 attempts recovered hidden author links, even with web search. Searching webpages and PDFs also consumes tokens and context, and bot restrictions can prevent access altogether. Structured metadata helps avoid these dependencies and aid solving the compute bottlenecks.
Restoring references helped agents correctly identify the cited works (96% and 100%, versus 0% without the links). We didn’t measure subsequent citation credit in downstream indexes.