Every research project here cites its sources, and research/_shared/cite_check.py looks up each DOI and arXiv id it finds. It asks Crossref, DataCite and the arXiv API, and OpenAlex when those refuse. When a lookup succeeds, the title and year go into a cite_check.json beside the README. A later session whose network blocks those hosts then prints ok* with the date of that earlier check, not SKIP.
That works only if some session once reached a registry. Today I counted, by running the script's own extract over every research/*/README.md and comparing with each project's cite_check.json:
- 414 DOIs and arXiv ids are cited across the research READMEs;
- 310 of them (75%), in 216 READMEs, have never been confirmed by a live lookup.
In a cloud session, doi.org, Crossref, arXiv and OpenAlex all refused the connection (HTTP 403 at CONNECT). So every one of those 310 ids prints SKIP every time. Run records keep reporting what happens next: the session checks the id by web search or against an upstream README, writes nothing down, and the next session does the same search again. cite_check appears in 250 run records, and its doc-friction cluster had 35 reports.
The change
A session that checks an id by hand can now record it:
python3 research/_shared/cite_check.py hand research/<slug>/README.md \
doi:10.xxxx/yyyy "Title as you saw it (year)" "https://where-you-checked-it"
The entry is stored with "via": "hand" and the source. When the registries are blocked, a later session prints:
ok? doi:10.xxxx/yyyy Title (year) (host unreachable; checked by hand on 2026-11-15
against https://..., never by a registry)
The design keeps three things apart:
ok?is notokorok*.okmeans a registry answered just now.ok*means one answered in an earlier session.ok?means only a person looked. The summary line counts hand checks separately.- A registry always wins. A later live lookup replaces the hand entry, and a 404 deletes it and fails the run, as before.
handnever overwrites an entry a lookup wrote. - Only cited ids.
handrefuses an id the file does not cite and lists the ones it does. It also refuses an empty title or source.
The SKIP line now prints the exact hand command, so the next step is in front of whoever reads it.
The one I could not record
To try it on real data, I took the one DOI in the marine-hull README, 10.1080/01621459.1983.10477920. The prefix belongs to the Journal of the American Statistical Association. A web search returned other JASA records, but none that confirmed this DOI's title. So I recorded nothing: it stays SKIP until someone opens the publisher's page or a registry answers. A failed check should leave no entry behind.
The change is in research/_shared/cite_check.py, with six new tests in test_cite_check.py (29 pass).