How this site was made

The homepage says I direct teams of AI agents to produce client work, and that I have already done it. The nearest evidence is in front of you: the site you are reading was produced that way. This page is the record of how, held to the same standard as the scanner: publish the conduct, publish the failures.

The shape of a panel

A round starts with a written brief and a facts file that binds every agent: what may be claimed, what may not, and the client boundaries that survive any rewrite. Six agents then run concurrently, each in an isolated working copy of the site, each with one assigned lens, each blind to the other five. They are drawn from three model families made by three different companies, because agents from one family agree with each other too easily, and agreement was not the point.

Each agent returns a written report. I read the six against each other, verify any factual claim before acting on it, rule on the conflicts in writing and merge one result. Nothing runs unattended: every run happens inside a working session that I start, watch and end. The team is software; the accountability is one person.

The record

Three rounds, August 2026, six agents each.

Round one’s lenses: visual design; UX and information architecture; the first ten seconds; concept and proposition; the commercial path; and an adversarial line-by-line read. Round two, against round one’s output: buyer simulation; subtraction; rail architecture; objections; first-email mechanics; and identity, with a fresh adversarial pass. Round three re-angled the proposition and produced the page you are reading.

The same method produced two other kinds of work. The email campaigns for the business in the homepage findings were agent-built inside Klaviyo through the API; every send waits for the founder’s approval, and no campaign has been sent to a customer. The market research behind this firm’s own positioning was agent-produced with every number labelled verified, estimate or unverified, and the register counts checked against primary sources: the UK CAA’s approved-organisations list, ONS business counts, Companies House. That research is internal work, not client work, and the labels are the discipline: an estimate is never dressed as a count.

The disagreements are the point

If six agents agreed, the panel would be theatre. They did not, and the rulings are part of the record. Three examples, from the merge reports:

One agent argued the headline should lead with the transformation; another, timing a first impression, argued the existing headline should stay if a verifiable finding entered the first viewport. Ruled for the second, because its condition was met and an outcome-first headline loses the category faster than it gains the scroll.

One agent wanted the findings as collapsible rows; the design lens wanted compressed cards. Ruled for cards, against my own initial preference, because collapsed content does not print, and a report that hides its evidence on paper fails the house standard.

One agent, cutting the page for length, wanted to delete the sentence “The product named after the ingredient could not lawfully claim it.” as a restatement. An earlier pass had named it one of three untouchable sentences. The protect list won; the sentence is still on the homepage.

What the panels got wrong

The first round’s merged page shipped with three wrong sentences of its own. It promised that the free page read “confirms it is worth the fee”, a valuation the method refuses to give. It silently dropped two settled boundary sentences from the audience-recovery step: “I do not produce a report from this check”, and the requirement that the fee be justified in the client’s own campaign figures. The second round’s adversarial pass, run blind by a different model family, caught all three, and they were removed or restored. The errors leaned toward overclaiming, which is exactly why a later pass reads every earlier merge. The scanner’s own failure log records the same lesson from a different tool: its first list produced three findings, all three wrong, all three caught by hand before reaching anyone.

What stays private

The six raw reports behind each round are not published. They discuss a client engagement in detail and this firm’s own commercial research, and they were written for me, not for you. The rule is the same one that keeps the findings business unnamed: a client’s private weakness is not my marketing asset, and neither is the candour of a working document. If we are in a real conversation about real work, I will walk you through them on a screen share, the same way the audience-recovery check happens inside your own account.

What you can check, and what you cannot

You cannot verify from outside that the runs were blind, and I will not pretend otherwise. What you can check: this page against the standard it claims, no scripts, no tracking, a boundary beside every claim; the homepage’s evidence section, which states what is proven and what is not; and the offer above, which puts the unpublished record one conversation away. Producing pages proves production. It does not prove a customer bought anything, and no page of this site claims otherwise.

This record is current as of August 2026. If the site changes materially without a panel, this page will say so.