Match Evidence to Consequence

Practice. Choose one chapter or essay of about three thousand words and list every checkable factual claim. Mark each H, M or L for the harm if it were wrong. For every H, write what verification would require (which document, which second source, whom to ask), then do it, and mark the claim verified, qualified or cut, spending ninety minutes on the H claims alone. You end up with a ledger and at least one sentence in the draft that changed. Each H claim you could not verify is a place where the draft has been asking readers to take your word.

Before any checking, ask two questions about each claim. What happens if this is wrong, and to whom? And how thin is what I actually have? For a decade, organizations urged clinicians to encourage postmenopausal women to take hormone replacement therapy, on the belief that it substantially reduced the risk of heart disease. Looking back in 2008, the authors behind the GRADE system for rating evidence pointed out that the belief had rested on observational studies with inconsistent results, which under their scheme count as very low quality. Randomized trials later found that the therapy fails to reduce cardiovascular risk and may even increase it.

Illustrative claims, like the weather in an anecdote or the name of a street, can rest on a secondhand source if you label them as such (Practice 71). Load-bearing claims, such as a statistic that anchors an essay or a date that organizes a narrative, need a primary source or two independent ones. Claims that could injure a person (an accusation, a diagnosis, a financial or legal assertion) need documents, independent corroboration and the person’s chance to respond. A claim’s tier depends on the weight it carries in your book, and the decorative statistic on which three later paragraphs quietly build has already been promoted without your noticing. When the top tier cannot be reached because the records are sealed or the only witness is dead, the claim is cut or stated as exactly what it is (Practice 78). It does not get inflated to fit the paragraph.

Verification is labor, labor is finite, and errors do not cost everyone the same, so the checks go where the damage would land. A language model can deliver the trivial claim and the ruinous one in the same calm register. Its tone tells you nothing about which tier a claim belongs in.