AI in production
Shipped
Designed and built by me · Horizon Web
Agents that review every site before launch
Before a site goes live and reaches a payment provider, it has to pass a review on four rule sets: accessibility, consumer law, cookie consent, card scheme rules. I designed and built the agents that run it. The design work was not the automation. It was the output contract: what the agent may claim, what it has to prove, and what a person does with the result.
- 4rule sets reviewed on every launch
- 0findings without evidence: they are not written
- 32live sites under the same review
01
Four rules
Every finding cites its evidence: a URL, a file and line, or a screenshot. What the agent cannot verify from outside never passes as compliant; unknown is a state, not a pass. A missing fact comes back marked as blocking, never filled in with something plausible. And the report is not allowed to conclude that a site is compliant: it counts what holds and what still blocks, and the person decides.
Blocking · subscription terms
The payment page does not state the monthly price that follows the trial.
evidence: /checkout · screenshot 14:02 · rule: consumer law, pre-contract information
What holds: 11 checks. What still blocks: 2. The report does not conclude.
A finding as the agents write it: the rule, the evidence, what holds, what blocks, and the person's three ways out.
02
Why it is designed this way
A review that reads well and cannot be checked is worse than no review: people stop using a reviewer they cannot verify. So most of the effort went into making every line checkable, the finding, the rule it breaks, the proof, the fix. Fluency was removed on purpose. Remediation is a list of blocking items with their evidence, not a verdict to argue with, and accessibility is part of the review rather than a pass at the end.