An unsourced statistics checker catches the number nobody can find.

A paragraph with a flawed argument invites argument. A paragraph with a percentage invites agreement, because checking it requires leaving the document.

Reviewers under time pressure triage by plausibility, and generated statistics are plausible by construction: a growth rate in a range the industry reports, attached to a claim the reader already believes.

The failure is invisible until somebody with a stake in the number asks where it came from, which is often a regulator, a competitor or a client.

The rule is deliberately blunt: a figure expressing magnitude without a visible source near it. It is designed to start a question rather than answer one.

The second failure is subtler than an invented figure: a real source, correctly linked, saying something adjacent to what the copy claims.

Generated drafts do this by summarising a plausible body of knowledge rather than a document. The link is real, the number exists somewhere in that source, and the sentence still overstates it.

So verification has two steps. Find the source, then read it and confirm it says what the sentence says, at the level of specificity the sentence implies.

Run the checker and get the list of figures. For each one, decide in a single pass: sourced and accurate, sourced but overstated, or unsourced.

Sourced and accurate needs a citation the reader can open. Overstated needs the sentence narrowed to what the source supports. Unsourced needs a source or a rewrite, and rewriting to a softer claim is usually faster than finding a study.

Then re-run to confirm the count dropped for the reason you think. A number that disappears because a paragraph was deleted is a legitimate fix; one that disappears because the sentence was rephrased into vagueness is not.

In health, finance, employment and consumer protection, an unsupported statistic is not an editorial problem. It is a statement a regulator can read literally, with a complaint process attached.

Generated copy has no instinct for that boundary, so the gate has to. Any claim about outcomes, efficacy, safety, savings or performance should require a named approver, regardless of how the sentence is hedged.

The cheapest control is a list of forbidden claim shapes for the pipeline: guaranteed, clinically proven, risk-free, number-one, and the percentage that implies any of them without a footnote.

Sometimes a figure is right and the source has vanished: a paywalled report, a company page that has since changed, an internal dataset that cannot be published.

Two honest options exist. Cite the internal source as such, with a date, and accept that it is not independently checkable. Or remove the number and describe the direction instead, which is usually enough for the reader.

What must not happen is presenting an unverifiable figure in the same voice as a verified one. That is the specific failure the checker exists to prevent, and it is the one that damages a brand when a journalist asks.