Paper 01 · 28 October 2026

One Conductor, Many Instruments

How we run one orchestrating model, outside models commissioned per question, and a separate checker on our own work ... with the owner holding every key.

Every lab now ships a capable model. The question we kept running into was not which one is best. It was how to put several of them on the same job without anyone, human or machine, losing track of who decided what.

This paper describes the setup we use on our own work: building and checking the websites for Paul Ann Labs, Paul Ann Media and HomeTimeJobs. It is not a client deployment. It is small, it is young, and it has already failed in ways we describe below.

The seats

People will say the setup is controlled by AI. It isn’t. It is run by a model and governed by the owner. The difference is the whole design.

The seat is the contract, not the model

We call the outside voices interchangeable because the rules do not change with the model in a seat. The brief is written once and sent unchanged to each. When Gemini failed twice on a tall review image, we sent it a copy 1,000 pixels wide and the run went on. Changing a voice changes what we hear. It does not change how the work gets checked.

There is one more rule about outside voices. When a question also goes to the checker, the checker sees what the outside models said before it concludes, and answers point by point: agree, disagree, or measured differently. Their text can never widen a ticket or change a rule.

The fence

• One shared folder, two inboxes: tickets go to the checker, results come back.
• The checker acts only on tickets marked open. Draft means waiting for the owner. Closed means done.
• A stop file. While it exists, the checker and every helper do nothing.
• Nothing is overwritten. A correction is a new file.
• Every claim in a result carries its basis: measured this run, read from a file, or assessed.
• Its rules bar the checker from pushing, sending, posting, buying or signing, and from holding any payment or personal social account credential.
• Trust comes in phases. The checker’s first phase is watching: daily site checks and expiry dates, no new access. A building phase is decided only after two clean weeks.

A worked example: the rollout review

On 27 October, after a round of new home pages, we asked three outside models one question about all three sites: does the first screen make the right visitor stay? Scores, 1 to 10:

Then the checker tested the six claims that could be checked, on the live sites. It agreed with one, disagreed with three, and found two that held only in part or with different numbers. One model said our phone photo cut a truck in half; the live file showed the whole pickup and both workers. The same model said a teal button failed contrast. The checker measured 4.34 to 1, under the 4.5 to 1 line. That one was right.

All three voices agreed that on a phone, the button belonged on the first screen. The checker put numbers on it. On a 375 by 667 screen the buttons ended at 752, 791 and 836 pixels ... below the fold on all three sites. We shortened the phone photos, moved the button above the body copy, added a phone menu and adjusted two accent colors until text on them passed. After the owner pushed, the checker measured again: the buttons now end at 584, 595 and 640 pixels, inside the 667-pixel screen.

The recheck covered 14 items and all 14 passed, with a few small leftovers named. Its contrast audit of text on solid backgrounds found 0 failures across 54 page and width combinations.

What failed

The orchestrator’s own check of the new photos did not test the first screen on a small phone. The owner found it, looking at the sites on a phone. Every hero review now includes a 375 by 667 check.

The orchestrator edited a site while the owner was pushing. The live commit pointed at an image that had not been committed; a later commit fixed it. The rule now: no edits between handing over a push and hearing “pushed.”

Generated pages once showed an internal draft note on the page itself. It was caught before it went live. A browser test also ran into sites the browser tool could not reach, three times, before the orchestrator handed it to the owner.

The orchestrator also runs internal reviewer roles. They are the same model wearing a different hat. They are useful, and they are not independent. That is why the checker and the outside voices exist, and why the owner still reads everything.

Still open: one navigation link on HomeTimeJobs is 23 pixels wide, a pixel under the minimum. It is on the list.

What we don’t claim

• It is not autonomous. It runs on the owner’s approvals.
• We have run it on our own sites only. This paper is not evidence of results for anyone else.
• We have not measured which seat adds the most. When we do, we will say how we counted.