VenkyKash’s jev-use-exp lets Jev choose actions from an application-built menu, with Cua Driver handling observation and execution. Its desktop experiment records failed tasks and passes the observed screen text back for review.

Keep selection, execution, and verification separate

  • The TypeSafe adapter sends the goal, current observations, history, and bounded action choices.
  • The native runner checks the selected ID, confidence, and probability distribution before resolving an action.
  • Exact completion checks require a unique matching native element and the expected value in a complete observation.
  • A review handoff carries source identity, text, and completeness information with independently_verified: false.

The calling assistant reviews that evidence after the loop stops. The handoff is a data record; it does not itself launch another model or finish the task.

Calculator exposed state-management failures

The repository’s September 19 report describes a multiplication retry that cleared only the current operand. A leftover expression survived, and the subsequent inputs produced the wrong result.

Other reported Calculator runs stopped on partial expressions after selected actions failed the configured confidence cutoff. TextEdit was blocked by ambiguous windows before typing. Those outcomes stayed labeled failed, incomplete, or blocked despite passing software tests.

These are the project’s historical reports. The raw run logs are ignored rather than committed, and the desktop outcomes were not reproduced here.

Historical safeguards differ from today’s defaults

The failure report records decisions rejected under an earlier confidence policy. The inspected native runner now defaults to 0.1, which its documentation calls a requested trial rather than a calibrated safety guarantee. Its cutoff tests confirm that current behavior.

A lower threshold may allow more actions through. It supplies no evidence that the remaining actions are correct. Typed choices constrain the action menu; representative tasks still have to establish decision quality.

The browser research report also records an external link being followed before the host check stopped the run. Its permitted-host list should therefore be treated as a post-navigation check, not a network sandbox.

Offline checks are reproducible

At commit 385a91cfc2db039cf267fe67b421cb3f2b7ffa9f, a local Python unittest run completed 177 tests, with two skipped and no failures. The initial focused invocations needed PYTHONPATH=python; the full rerun used that setting.

Those checks exercise software behavior with fixtures and mocks. They do not establish live Jev accuracy, successful desktop control, or billing attribution. The project separately documents an unresolved mismatch between client-recorded API responses and dashboard usage.

Caveat: Python tests were run; live inference, desktop actions, and TypeScript checks were not. This was not a security audit.