Keep the Writer Away From the Judge

AI agents handle much of my work across many client projects, with me directing them. I built HALO to manage those projects. Its central rule is that the model writing words for publication never evaluates them.

I made that split after news that one AI vendor marks the text its models write. A writer grading its own work is like a student grading their own exam. The student has a reason to see the answers as correct. A separate reader has a different job.

How does HALO keep writing and judging separate?

Independent hosts run a separate writing model that creates every published sentence. Reasoning models give the writer its brief, assess the page and share their findings. They do not write copy. Writing and review stay separate.

The division gives each model a defined task. A reasoning model can point to a problem and say what needs to change. It cannot rewrite the passage itself. That instruction keeps the review from quietly turning into a second draft written by the judge.

Before a writer starts, I give it my style rules. A writer that can't receive those rules won't run. I also rotate passages among multiple writing models, so no single model writes a long, uninterrupted stretch of a page. That lets me split up the writing while keeping the judge from taking over.

What does the check review?

The check looks for passages the assigned writer did not produce and lists them. It also opens every link and checks quotations and contact details against the brief. If it cannot verify an item, it reports "could not verify" rather than calling it fine.

That wording matters because an unchecked claim can otherwise look settled. A report that says a quotation or contact detail could not be verified tells me where to review; it does not turn uncertainty into approval. The check has a clear boundary: it reports what it could confirm and flags what it could not.

Before I trust a check that reports zero problems, I plant a problem and see whether it catches it. A clean report without that test gives me no evidence that the check is working. The test is about the checker, not the page. If the check misses the planted problem, I do not treat its clean result as reliable.

How does the judge grade a page?

For each page, the judge scores eight measures, including experience, expertise, trustworthiness and readability, against a minimum I set. I use the middle score from three separate reviews because a single review can shift by up to ten points on an unchanged page.

Using the middle gives me a score less exposed to one unusually high or low read. It does not make the page perfect or prove every judgment is right. It gives the review a consistent way to compare the page with the floor, while the separate checks handle details such as links and quotations.

When the judge finds a problem, it describes the change. It never writes the replacement sentence. That instruction sends the fix back to the writer, where the copy belongs. The judge can assess a page and explain what failed, but it cannot become the author of the repair.

Why keep one model from doing every job?

A model that writes and grades the same words has both roles in one place. Separating them lets the writer focus on the brief and style rules while another system looks for problems. Different models handle different tasks, so one model does not have to write, grade and revise the same copy.

That separation does not make every report correct. The planted-problem test checks whether the checker can catch a known issue, and the three reads reduce the effect of one unusual grade. Each method has a specific purpose, with a limit: a check reports what it cannot verify, and a judge describes a fix without writing it.

Common questions