Catching a summary of an abstract the model never read past

1 · Learn the move · Fact-check list

The most dangerous paper summary is the fluent one built from an abstract plus the model's expectations of what such a paper usually says. It arrives with methods details, effect sizes, caveats - some not in the paper. The fact-check list turns suspicion into procedure. For every claim: which sentence of the supplied text supports it - quoted, not gestured at. Claims with no source sentence become UNVERIFIED - text does not support. Two tells, every time: specificity beyond the source - an n, a percentage the abstract never states - and section claims, methods or limitations when only an abstract was supplied. The prevention is cheap: every summary opens by declaring what text built it. Abstract-only summaries that say so are honest tools; abstract-only summaries wearing full-text confidence are how a wrong claim gets a citation.

Fact-check this summary against the text it claims to summarize. For each claim: (1) the claim, restated in one line; (2) the verbatim source sentence from the supplied text, or NONE; (3) verdict: SUPPORTED / UNVERIFIED - text does not support / CONTRADICTED. Run the two standing tells afterward: list every number in the summary absent from the source, and every claim about a section (methods, limitations, statistics) the supplied text does not contain. End with the corrected summary, which must open: "Based on [what was actually supplied]."
[summary] [source text]