Can Jev work with screenshots and images?
Verdict for the sample idea
Run it on
a reasoning LLM
Jev cannot take images, so this goes to a model that can.
Sample idea
Look at a screenshot of our app after each build and decide if the layout is broken: overlapping text, cut-off buttons or missing images.
Explanation
Jev does not accept images today. The input is text or structured data. A task that starts with "look at" is not a Jev task, however well the rest of it fits.
This page exists because the rest fits so well. "Is the layout broken?" is a yes or no. It runs after every build. It needs a probability. Every signal says Jev, except the one that decides. The first version of jev-fit made exactly this mistake and recommended Jev for this idea. A separate question about image input had fired correctly. The verdict question never saw that answer, because questions in one call run independently. The fix was a veto in code. If an idea needs images, audio or video, Jev is removed from the verdict. I describe the full lesson on the report page.
You have three options for the task itself.
Use a vision model. A multimodal LLM can look at the screenshot. It is slow and costs more, and for a check after each build that is acceptable.
Use plain code. Snapshot testing compares a new screenshot with a reference image, pixel by pixel. For "did the layout change?" it is exact, free and already exists for iOS, Android and the web. It answers a narrower question, and often that question is enough.
Convert the screen to text, and then use Jev. The accessibility tree gives you every element with its frame and its label. Code can detect overlapping frames and truncated labels with geometry. Jev can then judge what geometry cannot: is this label sensible for this button? Is an error message on the screen?
The common trap is OCR as a shortcut. Text from OCR has lost the layout, and the layout was the question.
Related tasks
Verdict produced by Jev on 2026-09-19, rubric version 2. Text written by a person. How the tool works: the report.