An AI writing quality check needs a rubric, not a vibe.

Score in this order, because a failure at the top makes the rest irrelevant.

Accuracy is partly mechanical: statistics without visible sources, absolute claims, quotations without attribution. The rest needs a human who can open the source.

Structure is largely mechanical: heading ranks, duplicates, a long piece with no sections, paragraphs that run past 220 words. Clarity is measurable enough to be useful: average sentence length, flat rhythm, jargon density, repeated openers.

Voice is the least mechanical and the most recognisable. Stock phrasing, hedging and symmetrical triads are the visible surface of a draft that was assembled rather than written.

Averaging dimensions produces the same problem as averaging findings: one unsourced regulatory claim can vanish behind four clean sections.

So report the dimension scores and the blocking findings separately, and let the blocking list decide. Accuracy failures should never be averaged into a total.

The same caution applies to improvement over time. A rising score sometimes means the rule was removed rather than the writing fixed; compare the findings, not the number.

Generated drafts are rarely incomplete in obvious ways. They are incomplete in the way that survives a read: a heading promised a section that is a paragraph, a process described without its prerequisite, a comparison without the alternative the reader will ask about.

The check that matters here is the brief. Compare the draft against the original ask, line by line, and mark what is missing. Ten minutes of that catches more than a full line edit.

For procedural writing, completeness is testability: can a reader who matches the stated prerequisites follow the steps and arrive at the described state.

A landing page and a policy document do not share thresholds. The first can tolerate a short paragraph with a bold claim; the second cannot.

Set expectations per channel: what accuracy evidence is required, what structure is mandatory, what voice tolerance applies. Then the same report can serve every channel without pretending the standards are identical.

Write the thresholds down next to the brief, so a writer knows what a pass looks like before drafting.

Anywhere accuracy carries consequence, the rubric ends in a named decision: an editor, a compliance reader, a subject specialist, depending on the claim.

That signature should state the scope. Reviewed for factual accuracy in the pricing and eligibility sections is auditable. Approved is not.

The report and the signature together give you the two things a quality process is for: repeatability and accountability.