What it cannot do

Quotes: a quote that is really there can still be out of context ("is available" cut from "is available only to staff"). That is for a person to judge. An elided quote ("A … B") is not matched.

Figures: it checks that a number was supplied, not what it is about: "40 women" in your materials supports "40 women" and also "40 men". It does not catch an invented name, only numbers. Currency symbols are not compared.

Citations:

• It proves a source exists, not that it says what the sentence claims. A verified citation can still be attached to a false claim, and a fabricated one can be written to match a real record. The messages say "a record exists", not "this is correct".
• Coverage has gaps. Books, reports, chapters, grey literature and non-English work are thin in these indexes. A real work can come back failed (a real WHO report did, in the test corpus). That is why the wording is "not verified", and why strip() is opt-in.
• Direct links cannot be checked from a browser (most sites do not allow it). They are reported as check.
• approximate citations name an area, not a paper. Assay can only say whether the author has related work.
• The sources limit free use. Without mailto and an OpenAlex key, a busy app will be throttled, and the result is then unavailable, not failed.
• The test corpus is 30 citations, hand-labelled, from one afternoon with free models. npm run corpus runs it against the live sources. It shows a change moving the numbers; it is not a benchmark, and no accuracy figure should be quoted from it.

Privacy

Only the author names, title, year and link of each citation are sent, never the surrounding text. They go to: CrossRef, OpenAlex, Open Library, doi.org, DataCite and PubMed (the last three only when a link calls for them). Nothing is sent anywhere else, and nothing is logged.