Benchmarking Jev's Box Tool on ScreenSpot (Arthur Katcher)
External benchmark write-up by Arthur Katcher (2026-10-17): Jev 1.14 box on ScreenSpot-v2 and ScreenSpot-Pro with official scoring, plus a zoomed second look and five harness rules for click agents.
What this guide points to
Arthur Katcher published I benchmarked Jev’s box tool on ScreenSpot-v2 and ScreenSpot-Pro (2026-10-17). He sends one screenshot and one box question per click, clicks the centre of the returned box, and scores with each benchmark’s official eval code.
Reported results: 84.36% on ScreenSpot-v2 and 18.66% on ScreenSpot-Pro raw. A second look on a half-size crop around the first box lifts Pro to 42.76%; a post-hoc third look reaches 60.15%. The article explains why target size in the scaled image drives most of the gap, and why zoom does not help on phone screens.
This page is an unofficial field manual pointer — verify numbers against the article, the jev-click-bench repo and current TypeSafe docs. Box is a preview feature.
Takeaways
- Ask one box question per screenshot and click the box centre.
- Add a zoomed second look on desktop and web screens; skip it on phones.
- Act on zoomed answers at high confidence only; send the rest to a check or a stronger model.