The Yardstick Problem: When an LLM Fact-Checker Flags Most of What It Reads

I build a pipeline that drafts illustrated companion chapters for Indian primary schools and checks every statement in them against the NCERT textbooks before an editor reads the chapter. On a Class 3 chapter about water, the checker sent 50 of 87 statements to the editor. I took this to mean the chapter was badly sourced. When I sorted the flags by cause, most turned out to come from the checker. It held every statement to one test, a verified quotation from a source passage, and that test does not fit most of what a children's chapter contains: definitions of everyday words (NCERT rarely defines them), activity steps (which are neither true nor false), lines of story, and supported facts that a safety rule sent to a person anyway. I changed the checker so that each kind of statement gets a test that suits it, and enforced one rule in code: a statement of fact can only be cleared by a quotation. In a controlled comparison on two drafts of the chapter, run twice each, the mean number of statements sent to the editor fell from 46.8 to 16.0. On a copy of the clean draft carrying one planted error of each kind (a false definition, an unsafe instruction, false claims in dialogue and in narration, and a plain false fact), and on an older draft with four planted errors, both versions caught every plant in every run. About half of the reduction came from definitions and instructions now judged by a second model pass, and those judgments were not audited. Two raters audited the facts that left the list: none was inaccurate, but two of seven newly confirmed facts were only partly supported by the quotation, and the old checker had been right to flag them.