What it is
The first public GUI-grounding numbers for Jev's box tool.
What it does: Runs official ScreenSpot evals and tests zoom-in second and third looks.
Why it belongs in the showcase: Reproducible logs, pre-registered prediction, and five concrete harness rules.
Caveats: Preview feature; single model version.
Evidence
Public MIT repo (0★ at intake) created 2026-10-17 with harness, logs and audits for 2,853 clicks. README table: v2 84.36% raw / 87.89% with 2x zoom; Pro 18.66% raw / 42.76% with 2x zoom / 60.15% with a post-hoc 4x step whose prediction was committed before it ran. Costs per 1,000 clicks $0.15–$0.48 as billed by OpenRouter. Companion write-up on arthurkatcher.com and DEV.
Caveats
- One model version (jev-1.14-20261014); Box is a preview feature
- Author-run benchmark; other models' scores are as published
- Still screenshots only; real agents must also detect missing targets
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.