Verification before publishing is a sequence, and the order is the whole trick.

Run the draft through the checks before anyone edits it. Placeholders, unsourced figures, absolute claims, broken links, missing alt text, heading structure, stock phrasing.

Doing this first means the editorial pass is not interrupted by a TODO left in paragraph four, and the writer sees the whole defect list while the draft is still cheap to change.

Fix mechanical findings in the same sitting. They are almost all one-line changes, and leaving them open means they reappear in review.

List every assertion in the draft: numbers, dates, quotes, comparisons, superlatives, promises about outcomes. For each, either find a source a reader can open or narrow the sentence.

The order matters. Source first, then rewrite. Rewriting a number into vagueness removes the finding and keeps the misinformation, which is worse than the original because it is harder to challenge.

Anything that survives this pass with a source gets the source recorded next to it. That list is the artefact a reviewer or a regulator will ask for.

Read the headings alone. Each should promise something the section delivers. If a heading could sit on any page in the category, the section underneath it is filler.

Then read the first paragraph of each section in sequence. If they all share a shape, the piece is being carried by its template rather than its content.

This is also where length gets honest. Generated drafts are usually too long by a third, and the surplus is in the transitions and the recaps.

Cut one sentence per section to four words. Let another run long. Delete the sentence announcing what the next one will say.

Replace abstract nouns with the thing they describe, and check the tone against the moment. A cheerful register on a sensitive subject is a defect that no rule detects.

Re-run the style checks afterwards. The second measurement is the one worth recording, because it shows the draft was edited rather than generated.

Map the findings to decisions agreed in advance. Unsourced statistics in regulated copy block. Missing alt text warns. A placeholder blocks, always, because publishing one is a visible error.

Where the copy carries consequence, a named reader signs a scope: reviewed claims and eligibility wording in the pricing section, for example. Not approved, which says nothing.

Record the report, the decision and the signature with the piece. That record is the difference between a process and a claim about a process.

Sample published pages monthly. Gates measure what they test, and the defects that escape are the ones nobody wrote a rule for.

Watch for the two drift patterns: a template change that strips metadata or alt text across hundreds of pages, and a factual claim that was true when written and has expired.

Feed both back into the rule set and the brief. A publish gate that never changes is one that has stopped learning from what actually ships.

Most of what a publish gate catches was invited by the brief. A prompt that asks for a persuasive page will produce confident claims; one that names the audience, the evidence available and the claims that are off limits produces something that passes.

So the cheapest control sits before drafting: what this page must prove, what it may not assert, which figures are approved and who verifies anything new. Two lines in the brief routinely remove half the findings.

The piece you verified is rarely the piece that ships. The same claims get shortened for a social post, tightened for an email subject and compressed into an ad headline, and the compression is where qualifications disappear.

Check the variants as their own artefacts: a headline that drops the qualifier changes the claim, and a subject line has no room for the source note the page carried. Any figure that appears in a variant needs the same sourcing as the page.

Translation moves claims as well as words. A hedge becomes an assertion, a comparative becomes a superlative, and a quantity expressed in one unit arrives in another with a rounding that changes the meaning.

Route any translated page that carries regulated claims back through the claim audit, in the target language, with a reader who knows the market. Automated checks run on the source will not see the drift.

It happens constantly, and the resolution should not depend on who is senior in the room. Ask which rule fired, what it protects against and whether this case is genuinely different.

If it is, record the exemption with your reason so the next person inherits the decision rather than relitigating it. If it is not, fix the sentence. Suppressing a rule to meet a deadline is the decision that produces the incident six weeks later.

Track escaped errors: claims corrected after publication, placeholders found by a reader, links that shipped dead. Each is a rule you did not have or a pass somebody skipped.

Then look at override volume. A gate with a high override rate is not being obeyed, and a gate nobody obeys is worse than none because it produces a record of assurance that is not true.