An AI content verification workflow with the authority to delete.
Two-agent content production: one writes, one audits with the authority to correct and delete. 1,374 corrections on 100 articles is what that authority looks like in practice.
A model that writes will also invent.
A language model writes fluent copy at a rate no studio can match, and some of that copy is wrong. Not badly written; wrong. A figure that was never measured, a feature the platform does not have, a quotation nobody said. The prose is confident either way, which is the problem. Fluency is not evidence.
The usual answer is a person reading everything. That works at ten articles and fails at a hundred, because reading a confident paragraph for errors is the task people are worst at. The eye slides over what sounds right.
The playbook’s answer is not to trust the writer. It is to give a second agent the job of distrusting it, and the authority to act on what it finds.
Fluency is not evidence. A source is.
One writes. One audits, and can delete.
The writer drafts against a brief
The first agent writes the article from a brief that names the query the page owns, the silo it lives in and the claims it may make. It writes the draft. It does not publish it.
The auditor checks every claim
The second agent reads the draft as an adversary. Every figure, every statement about what a product does, every quotation is a claim to be sourced. It has no stake in the draft and no instruction to be kind.
Correct what can be fixed. Delete what cannot.
Where a claim is wrong but a sourced version exists, the auditor corrects it. Where no source can be found, the auditor deletes the claim. Not softens, not flags. A sentence that cannot be backed does not appear.
Count it, then publish the count
Every correction and every deletion is logged. The totals are the receipts: the site says how many times the auditor acted, and the log is what makes the number a fact rather than a boast.
The order matters. Verification after writing, by a different agent, with power over the text. Verification by the same model that wrote the draft is a model agreeing with itself.
What authority means: 1,374 corrections, 308 deletions.
The playbook’s reference build is 100 articles, shipped and inspectable. The writer drafted all of them. The auditor applied 1,374 corrections across the set and removed 308 claims that could not be sourced: about fourteen corrections and three deletions per article, from a writer that was already producing confident prose.
Authority is the word that matters. An auditor that can only flag produces a list somebody may read. An auditor that can delete produces a text with nothing unsourced left in it. The difference shows in the finished site: a reader who checks a claim finds a source, because every claim without one was already removed.
Those deletions are not lost work. They are the arithmetic of the release gate applied to prose: a check that refuses is worth more than a check that warns. And the audit log answers the Data question for content: the record of what was said and what was changed lives in your repository, not in a vendor’s dashboard.
The auditor can delete. So nothing unsourced survives.
Verified content is one thing. A governed agent is another.
Write, then verify sits after one query per page in the sequence and produces the pages the gate checks. The last system, the governed agent, applies the same distrust to an AI that talks to your client’s visitors.
Read the sharpest chapter first. It costs an email.
We’ll send the build-gate chapter in full: the checks, the code, and the reasoning. Read it. If it doesn’t change how you think about shipping client sites, you’ve lost ten minutes and kept a working release gate.
That did not send. Check the address and try again.