How Maark reviews an article before you ever see it

The uncomfortable truth about AI-written content is that the model that wrote it is a poor judge of it. Ask a language model to grade its own draft and it will mark generously — same instincts, same blind spots, same confident tone about a fact it invented. If you want a review that means something, the reviewer cannot be the author.

That principle is the spine of Maark's editorial gate: the stage that reads a draft after it is written and before it reaches you. Here is what it actually does.

A different family reviews than wrote

Maark's articles are drafted by one model family and reviewed by another. The writer runs on Anthropic models; the review panel runs on a different family entirely, and the writing family is structurally excluded from judging its own work. This is not a stylistic preference — same-family judging measurably inflates scores, so the reviewer is deliberately a stranger to the draft.

The review is a panel, not a single opinion. Each of several criteria — grammar and mechanics, clarity and flow, brand voice, usefulness and depth, and whether the piece actually matches the search intent it was written for — is judged independently. A strict first seat and an independent second seat weigh in; when they split, a third opinion breaks the tie by majority. And a reviewer that fails a draft has to say why and how to fix it — a verdict with no actionable instruction is thrown out as unusable, because a rejection you cannot act on is just noise.

Claims get checked against sources

Separately from the prose review, Maark checks the draft's claims. Statements are verified against the material behind them, and anything the draft cannot support is flagged rather than waved through. The writing stage is built to leave gaps honest: where a claim needs evidence you have not provided, the piece is instructed to leave it out rather than invent it. Maark will not fabricate testimonials, metrics, pricing, or customer logos to fill a paragraph.

There is also a deterministic verification pass that has nothing to do with taste: every link is checked over HTTP so a draft cannot ship a dead URL, headings and tables are checked for structure, and images are checked for real, non-generic alt text. These are the checks a careful human editor would do and usually does not have time for.

What happens when a draft does not pass

This is the part worth being precise about, because it is where honesty and marketing usually part ways.

The review can hold a draft and route it back for repair. A held draft does not quietly appear in your queue as ready-to-publish — it goes through a repair loop first, with the specific defects handed back to a revision stage, and the repair budget is capped so a piece cannot ping-pong forever. If it cannot be repaired within budget, it parks as blocked with an exception for a human to look at, rather than degrading into something publishable.

We will not overstate how universal that enforcement is. The review runs in a shadow-by-default posture: for a given project it records its verdicts and surfaces them without hard-blocking, and hard enforcement is switched on per project and per plan. We built it that way on purpose — you get the review's judgment visibly before you hand it the authority to stop a publish, not after. What is true in every configuration is the thing that matters most: a finished draft never publishes itself. It lands in your review queue, where you publish, preview, or skip. A person is always the last step before anything goes live.

Why a gate instead of a better prompt

You could try to prevent every problem with a cleverer writing prompt. We think that is the wrong bet. Prompts drift, models change, and the failure you did not anticipate is the one that ships. A gate is a second, independent system whose only job is to catch what the first one missed — and to leave a record when it does. It is the difference between hoping the draft is good and having something check.

The content pipeline help articles walk through the stages in customer terms, and why we publish review-first explains the philosophy behind keeping a human at the end.