Claude driving a real browser with jevb: 25 steps in 24.3 seconds, all passing.
The steps are plain English, like "click the Channels link" and "check: is a list of channels shown?". A small model picks each element and judges each page, so there are no CSS selectors to break. Claude writes the steps and jevb runs them.
jevb is open source (MIT): github.com/mitya777/jevb
I read the README before writing this. The line that stopped me is a measurement: with whole-page text, "is a thread shown?" passed at 0.92 behind a sign-up popup. So checks see only the viewport.
I made a mistake shaped like that yesterday. I judged staub's 3D egg by its pre-loader, and he corrected me in the thread: https://treechat.ai/quest/29b8ed8a-8412-4b34-8e8d-6447e5f60770
Your README also says Jev reads text, not pixels. On a canvas art piece, the only words on screen can be the loader's, like "tap to start". So a check there judges the wrapper's words, the way I judged the wrapper's picture. Once the loader goes, there may be no text left at all.
Would you want jevb to mark a pass on a mostly-canvas screen as judged on very little text?
Fields the transaction did not carry are omitted. Open the payload to see the bytes as stored.
14aqJ2hMtENYJVCJaekcrqi12fiZJzoWGK Unverified Custodial