research/_shared/cite_check.py looks up every DOI and arXiv id a document cites, through Crossref, DataCite, arXiv and OpenAlex, and prints the title the registry holds. That way a reader can compare it with what the text says the source says. Each live lookup is remembered in a cite_check.json beside the file. A later session whose proxy refuses the registries then prints ok* with the date the id was confirmed, not SKIP.
Two run records from 2026-11-16 reported where that broke down:
- Run on a blog post, it wrote
docs/posts/cite_check.jsonand told the run to commit it. No such file exists onmain, so the run could not tell whether posts should carry one, and left it out. - Given a project directory, it crashed with
IsADirectoryErrorand a traceback instead of finding the README.
What changed
A file outside research/<slug>/ no longer keeps a cache of its own. Instead it reads every project's cache: 21 projects have one on main today (git ls-files 'research/*/cite_check.json'). A post that cites a paper a research README has already confirmed prints ok* even in a blocked session, and the run finishes with "lookups not cached; nothing to commit". A hand check (cite_check.py hand) on a post is refused with a pointer to the project README that cites the same id. A post's hand check would otherwise land in the same untracked file.
A directory argument now checks the project's README.md, PROPOSAL.md and whitepaper, or every top-level .md when it has none of those. A path that does not exist exits 2 with "no such file". On research/codesign-cant the new form reads both its README and its proposal.
Checked on a real post, the 3D-printed robot one: its one DOI resolved live to Vision-controlled jetting for composite systems and robots (2023), and the working tree stayed clean. The checker's test file has three new cases and 32 tests pass.