What it is
An auditable harness for GUI grounding: given a screenshot and a goal, it returns a point and bounding box in raster coordinates. It uses recursive grid search and never clicks or captures the screen.
How it uses Jev
At each recursion layer, a vision model describes grid tiles goal-blindly. Jev receives the goal and neutral descriptions, returning a Choice distribution over tile IDs and NO_MATCH. Local code validates the distribution, applies confidence and margin gates, then recurses or refuses.
Primitives:choice
Technique worth stealing
Separate goal-blind visual description from goal-aware Jev choice over a closed candidate set.
Try it
pip install -e '.[dev]'; export AI_GATEWAY_API_KEY and JEV_GUI_VISION_MODEL; run jev-gui locate screenshot.png --goal '...'
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.92
- Shows a System One patternyes0.91
- Handles uncertaintyyes0.64
- Measuredno0.06
- Runnableunclear0.40
- Worth recommendingunclear0.47
- Model replicano0.06
- Problem scopescore on a 0–2 scale1.03
- About Jevyes0.98
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 22 Sept 2026.