# Claim-check worksheet Build n Bloom · v1 · September 2026 Run this before a sourced claim goes into a client deliverable. Worked example: the public-source review at `/demos/assessment/review-kit#public-proof`, also supplied as `public-source-review.md` and `public-source-review.html` beside this file. Nine questions. Answer them in writing next to the claim, not in your head. The worksheet checks whether a source supports a claim; it does not verify the source itself, and it does not replace specialist judgement about whether the claim matters to the client's decision. --- **1. The claim, as it will appear.** Write the exact sentence that will be read, including every number and qualifier. > **2. The source.** Title, publisher, publication date, and the date the underlying data was collected — they are usually different. Is this the primary record, or a summary of one? Note the exact location: paragraph, table or figure, and page. > **3. Snapshot or cohort?** Does the source follow the same subjects over time, or describe one moment? A claim about a success rate, a conversion, a failure rate or a change over time needs the same defined group observed at two points, with the outcome window stated. A single snapshot cannot supply that, however large it is. > **4. Denominator.** Who exactly is counted, how many, and who is not there — non-respondents, excluded categories, subjects that left before the count. State the response rate if there is one, and say what the count is a count *of*: people, organisations, cases, or documents. They are not interchangeable. > **5. Categories.** Are the categories mutually exclusive and exhaustive, or can one subject sit in several? If they overlap, does the source state the size of the overlap? Write down the definition of each category in the source's own terms, including any category that combines two different things (for example "piloting or planning"). > **6. The arithmetic, and whether it is justified.** Does the claim perform any calculation the source did not report? Set arithmetic on overlapping groups is often perfectly valid — if the source gives you the overlap, the parts can be recovered. What has to be justified is the *operation's meaning*: a ratio between two counts only expresses a rate if both counts describe the same subjects at two points in time. Write the operation, then write the sentence that makes it legitimate. If you cannot write that sentence, remove the calculation. > **7. Corroboration.** Is the second citation an independent instrument or dataset, or the same source republished — a press release, a summary article, a slide deck quoting the original? Two URLs are not two sources. > **8. The supported version.** Rewrite the claim using only what the source establishes, with its date, denominator and definitions attached. Expect it to be thinner and longer. Check that no word implies time passing, causation, cost or benefit unless the source measured it. > **9. The decision.** Keep as rewritten, soften, cut, or name the evidence that would resolve it — and who would have to produce that evidence, by when, for the decision the client is actually making. > --- ## Notes for the reviewer - Record the date you ran this and the version of the source you opened. Public documents can be revised or withdrawn; a locator without a date ages badly. - A missing record is not evidence that the thing did not happen. Say "not recorded in the supplied materials", not "did not occur". - If the claim survives, keep the answers with the working papers. The next reviewer should not have to redo the check to trust the sentence. Build n Bloom · · info@buildnbloom.io Free to copy, adapt and use inside your own method, with or without attribution.