The claim review editor: attaching a verdict to the sentence

A review note is not a review

The most common deliverable in an editorial review is a comment that says check the numbers. It is well meant and it fails for a structural reason: it moves the work to whoever reads the note, at a point when the draft is usually already scheduled. The person who could have checked the figure is the person who wrote the note, and they have moved on.

A verdict attached to the span that carries the claim changes the shape of the problem. Instead of an opinion about a document, you get a state that can be counted: this figure is verified, that one needs a source, this sentence is a hallucination. Countable states can be finished. Opinions about documents rarely are.

It also makes the review transferable. A second reader does not have to re-read everything evenly, because the draft now displays which claims somebody has ruled on and which are still open. That is the difference between a review that scales and one that is repeated.

Three verdicts and the commitment each one makes

Verified means you saw the evidence and it supports the sentence as written. Not a similar sentence, not the general idea: the sentence. If the source says revenue rose in the second half and the draft says it rose in the year, the honest verdict is not verified.

Needs source means the claim may well be true but nothing in the file, the briefing or your own knowledge backs it yet. It is a debt rather than a pass, and treating it as anything else is how unverified figures reach print with a clean-looking review attached.

Hallucination means the sentence contradicts the source, names something that does not exist, or states a figure nobody can trace. This verdict should be rare and it should never be softened because the surrounding paragraph is good. One fabricated citation poisons the credibility of the other twenty.

Unresolved is the fourth state, and it is the default. A mark nobody has ruled on stays flagged, so a ledger showing fourteen open claims is an honest picture of the work left rather than a false sense of completeness. That honesty is the whole value of the ledger.

• Verified: evidence seen, and it supports this sentence as written.
• Needs source: plausible, currently unsupported — a debt, not a pass.
• Hallucination: contradicted by the source, or untraceable.
• Unresolved: the default, and the only state that should never be exported as done.

What gets marked, and why that list is short

The scan targets four families: figures, percentages and money; dates and years; assertion vocabulary such as must, required by law, guaranteed, clinically proven and dosage; and attribution phrases like according to and research shows. Every one of those is a place where a document makes a debt to evidence.

Dates are included deliberately, even though most dates are correct. A date is a claim about time, and the failure mode — an out-of-date regulation, a superseded price, a fiscal year quoted as a calendar year — is expensive and invisible. Ruling on a correct date costs one click.

Opinions, definitions and ordinary narrative are left alone. The tool is not a fact-checker and it has no view on whether a claim is true; it finds the sentences that require a human decision and records what that human decided. That is the whole contract, and it is worth being clear about because the alternative — a tool that appears to check facts and does not — is worse than no tool.

In practice a thousand-word marketing page yields somewhere between ten and thirty marks. If it yields three, either the page makes no claims at all, which is its own finding, or the scan is being run on a section that only summarises.

A pass that ends with a number

Run find claims over the whole draft first and resist the urge to resolve as you go. You want the full ledger before you start, because the count tells you whether this is a thirty-minute pass or a day, and because a claim early in the piece often gets answered by a source you find while checking a later one.

Then sort work by risk. Legal, medical and financial assertions first: they carry the largest consequence and the smallest tolerance. Then figures, then dates, then attribution. Attribution last because it is usually the quickest — either the named person said the thing or they did not.

Resolve in batches. Open the sources you need, then rule on the claims that share a source while it is open. Context switching is the expensive part of claim work, not the reading.

Finish by reading the unresolved list aloud. Anything still flagged when you are tired is a claim you have decided not to check, and it should either be cut or explicitly noted as out of scope in the handoff.

Deciding what counts as a claim

The scanner looks for numbers, dates, assertion vocabulary and attribution phrases, and that covers most of what carries a debt to evidence. The gap is the sentence that asserts without any of those markers: the causal claim, the comparison, the implied recommendation.

Causes prevents and halves are claim verbs, and they usually attach to a noun that looks innocuous. The feature reduces churn by a third is a factual assertion with a figure, so it gets caught; the feature reduces churn is the same assertion with the number removed and it will pass every scan.

Superlatives are the other blind spot. The fastest in its class, the only platform that, the industry's first: each is a claim that can be disproved and each is written in the register of marketing, which is why it survives editing. Read the draft once asking only which sentences would be embarrassing if a competitor challenged them.

Opinion is not a claim, and marking it produces noise that teaches people to ignore the ledger. The right test is whether the sentence could be shown to be false. If it could, it needs a verdict, whatever vocabulary it happens to use.

Handing the ledger on

A ledger is only transferable if a second reader can understand the states without you in the room. Three verdicts plus unresolved is a small enough vocabulary to explain in a line, which is exactly why it works in a handoff note and why longer taxonomies do not.

Report the counts, not just the conclusion. Fourteen claims, eleven verified with sources, two needing a source and one hallucination tells a reader more than reviewed and approved does, and it tells them where to look first.

Name the unresolved ones explicitly rather than rolling them into a percentage. A reviewer inheriting the work needs the list, not the ratio, because they have to make a decision about each one and a decision needs the sentence.

And write down the calls you made in the other direction too: the claim you verified and then still cut, because it was true and unsupportable in the space available. Those are the hardest decisions in the pass and the easiest to forget.

A ledger for a recurring publication

For a weekly newsletter or a monthly report, the ledger compounds. Claims about the same figures recur, and a verified row from last month is a starting point rather than a fresh search, provided you re-open the source and check it still says the same thing.

Keep a standing list of the sources you trust for recurring topics, with the date each was last checked. That list is the most valuable document in a review workflow, because it turns verification from research into confirmation.

Retire rows that refer to a period that has passed. A verified claim about last quarter is not wrong, it is simply no longer a claim the next piece will make, and carrying stale rows makes the ledger look more complete than it is.

Short posts and long reports

A six-hundred-word blog post yields four or five claims and the pass takes ten minutes. A forty-page report yields over a hundred and the pass takes two days, which means the technique has to change: sample by section and escalate where the density is highest.

For long documents, do the sections a reader will quote from first: the executive summary, the findings, the recommendation. Those are the passages that travel without their context, and they are the passages where a missing qualifier does the most damage.

What the editor cannot do

It cannot verify anything. There is no network request, no source lookup and no internal knowledge; the three verdicts are records of decisions you made, not results the tool produced. A draft full of green marks is a draft somebody read carefully, assuming they did.

It cannot tell whether a source supports the sentence, only that you said it does. That subjectivity is the reason the tool exists: reviews are judgements, and a judgement recorded against a specific span is auditable in a way a general impression is not.

Nothing is saved. Marks live in the tab for the session, which keeps the tool honest about privacy and means you should copy the ledger before closing the page. For a review you need to keep, move the counts into the verification matrix, where each claim gets a row, a status and a source URL, and where the record persists on your device.