Evidence confidence
low. The implementation and one reported comparison motivate a test. They do not establish general cost or success rates.
What was observed
Choose a next Mac action from screen text and accessibility controls.
The author compares one decision for the same screenshot and goal; OCR and accessibility build Jev inputs.
Baseline
A frontier model reading the screenshot, with different preprocessing.
Finding
The report shows a much cheaper individual decision. It also exposes the work needed to supply structured screen state.
- Reported decision cost: about $0.0002 versus $0.032.
- Reported full step: about 1.5s with OCR versus 5.5s.
What the result does not establish
One decision does not measure complete-task reliability. Jev receives deterministic date processing; the comparator interprets pixels. Twelve-step costs are extrapolations.
What we would test in System One
Our inference: a desktop adapter needs a local observation layer. Direct image input to Jev is not demonstrated.
This recommendation is our interpretation of the study. Related research does not establish the quality of every Engine recipe.
Try a related workflow
Primary sources
Read this record in System One Bench. Source commits are pinned where available. Review dates describe our inspection, not the original run date.
Metric definitions and review method · Submit a correction or new result