Freeze the brief
Keep the user's original outcome and test flows. A vague requirement stays marked uncertain instead of quietly becoming a pass condition.
A framework for judging AI-made websites
“Done” is a claim.
What did someone actually see?
Evidence Loop keeps the person's original goal beside a tester's clicks and observations. It makes failures and uncertainty visible before another agent calls the work finished.
A working brief-and-report runner, not yet an autonomous browser-testing system.
The loop
One expectation. Multiple ways to be wrong.
Observe a fixed version
A tester tries the site as a person would: what they did, what happened, what evidence exists, and what they could not see.
Review, then decide
The report separates pass, failure, uncertainty, and untested work. A human decides whether to fix, clarify, or ship.
Try the idea
A sample report you can inspect
The downloadable runner generates a report from a frozen brief and observations. This sample is illustrative—no agent actually clicked this example site.
ILLUSTRATIVE · NOT A LIVE TESTexample.com
Goal: “A visitor can understand the page and follow a working link.”
PASS · read-pageOpened the example page and read its heading → saw a heading identifying an example domain.
NOT TESTED · follow-linkNo click was performed. The link cannot honestly be called working yet.
The package needs Node 18+ and a browser-capable AI app or human tester. It does not ask for your credentials.
Honest status
First milestone, not a universal tester.
Working now: a portable project brief, a Codex entry point, a runner that preserves observations, and a report that leaves untested work visible.
Not yet working: automatic agent orchestration, screen recording, browser adapters, automated fixing, or demonstrated portability across unrelated websites.
Next: test a real authorized website, independently inspect the report, then repeat on a different site without rewriting the core.