Every research README in this repository goes through research/_shared/cite_check.py before it lands. The script looks each DOI up in Crossref (then DataCite) and each arXiv id in the arXiv API, and prints the title and year the registry holds, so a mistyped or invented citation fails the run.
The trouble is the network in between. Agent sessions reach the web through a proxy whose policy differs from session to session, and some sessions are refused by Crossref, doi.org or arXiv. The checker already handled that honestly: a refusal prints SKIP, not FAIL, and an id confirmed by an earlier session prints ok* from a small cache. But a SKIP still means "nobody checked this", and the author has to open the source by hand.
How often? Counting the run records under prompts/runs/, 246 mention cite_check; refusals of arXiv come up in 27 of them, doi.org in 21 and Crossref in 18. In the last two weeks it was the fourth most repeated doc friction cluster (platform/ci/friction_rank.py --live --ungated).
The change. Before printing SKIP, the checker now asks OpenAlex, an open index that carries Crossref and DataCite DOIs and arXiv preprints (under their 10.48550/arXiv.<id> DOIs). A hit prints a new mark, ok~, with OpenAlex's title and year, and is cached like any other live lookup with "via": "openalex". Two rules keep it conservative:
- OpenAlex is only asked when the primary registry refused. A primary 404 still fails the run, because that registry is the authority on its own ids.
- OpenAlex not knowing an id is not a verdict, since its coverage is incomplete. That case still falls through to
ok*or SKIP.
Checked on real citations. With Crossref and arXiv forced to answer 403, the seven ids in the buildbot-sheathing README (five DOIs, including 10.1061/(ASCE)0733-9445(2001)127:4(433) with its brackets, and two arXiv ids) all resolved through OpenAlex with the right titles. One caveat shows up there too: OpenAlex merges an arXiv preprint with its journal version, so arXiv 2110.10107 reads 2022 rather than 2021. The output names the source so a reader knows which registry spoke.
Whether OpenAlex is reachable when Crossref is not is the open question; the sessions that probed it found it up. The next runs' records will say whether SKIPs drop.