CUA-JEV
Source code End-to-end Windows cases
Real model planning, Jev decisions, guarded execution, and task-level verification in recorded workflows.
REAL LAB · OPEN RESEARCH FRAMEWORK
One computer-use loop.
Many ways to act.
A text model interprets the live task and proposes grounded actions. The framework turns them into legal GUI and tool candidates. Jev chooses the next concrete action; guards and verifiers close the loop.
Windows implementation · recorded case studies, not a claim of arbitrary-task reliability
THE IDEA
Generalize the decision loop.
Keep execution grounded.
The model handles interpretation and candidate generation. Jev compares those candidates, including alternative routes for the same intent. The runtime constrains scope, executes the selected route, then checks the changed application state.
Structured observations are the reliable path today. Screenshot/VLM grounding is available experimentally, but visual-click success across varied applications has not yet been established.
WINDOWS · RECORDED WORKFLOWS
Watch the decisions unfold.
Each case below is backed by a successful local episode, its Jev trace, and a window-only recording. Counts describe that run, not an average or a benchmark.
Loading verified cases…
Recording and metrics notes
Action mix counts committed actions by execution channel. “Jev calls” counts policy decisions; “model calls” counts planner requests, with VLM calls shown separately. Recordings show only task windows; local path bars may be covered and a brief editor launch may be cut. Wall time comes from the raw run, not edited video length. One run per case does not establish a success rate or cost advantage.
RESEARCH STATUS
What the evidence does—and does not—show.
Visual grounding
VLM-derived targets are supported, but visual route reliability and benefit need broader live testing.
Broader generality
More unfamiliar tasks, applications, repeat trials, and macOS work after the Windows study.