Attest

AI agents open pull requests that look right, but someone still has to open the app and check that the bug is actually gone. I built Attest to do that step and leave proof either way. You approve a short requirement, like "press Escape once and the find box closes", and Attest freezes it. Then it runs the app before and after the change. Muse Spark 1.3 drives a real browser without ever seeing the code. A second model, Muse Spark Open, looks only at the screenshots and votes on its own. Plain code turns those votes into a verdict, and everything goes into a signed receipt with an evidence page: the requirement, every step with its screenshot, and each vote. Anyone can re-check all 220 published receipts without an API key. The demo is a real one. Devin's Board PR #913 said Escape now closes the find box. In the running app, Escape cleared the text but the box stayed open, so Attest rejected it. The follow-up, #914, passed. To be clear, this was a check after the fact on two stacked PRs, not a bug I found in the wild. It has limits, and I published them. On a held-out set it accepted 3 of 5 wrong fixes, because the requirements didn't cover the paths where those fixes broke. It never rejected a correct fix. In an exploratory follow-up I broadened the requirements to cover every path in each bug report, and it caught all 5. That came at a cost: 4 of the 10 runs on correct fixes didn't finish with a verdict, because of server errors or running out of steps. The lesson is simple: Attest checks what you ask it to check, nothing more. I built Attest during the hackathon. The open-source apps I tested it on already existed.