concepts
Concepts in 5 minutes.
UseDesktop evals are built around one loop: environment, task, model rollout, grader, and evidence.
Vocabulary
Environment
The resettable desktop world where the agent acts.
Task
The instruction and constraints the model must complete.
Rollout
One model attempt inside an environment.
Trace
The screenshots, actions, observations, and intermediate states from the rollout.
Grader
The program or verifier that scores the final result.
Evidence
The trace, score, verdict, failure notes, and audit context used to trust an eval result.
Desktop
The place to compare runs, review weak evals, and prepare training data.