ConsoleLog

Pixelsmith-Bench is an agent that fixes visual UI bugs by looking at a screenshot of the running page, plus a bench that measures whether the fix is real. Every coding agent says "fixed," and visual bugs are easy to fake: hide the element, nudge a neighbour, fix nothing. So the agent never grades itself. A hidden grader the model never sees checks each fix for correctness, cheating and collateral damage. On real bugs from Pico CSS the vision agent fixed 12/12; on real bugs the agent's own "fixed" claim was wrong 28% of the time (8 of 29); 140 cheat attempts against the grader got 0 accepted. Everything is open source, with confidence intervals and limitations in the README. Built this week: the agent, the 28-bug seeded bench, the grader and its tests, the SAM test, and the demo video. Pico CSS and daisyUI are third-party bug data (MIT).