Step 01
Enter the environment
Each challenge defines inputs, constraints, expected behavior, and evaluation criteria.
The learning loop
Versalist is built around environment, action, and reward.
Evaluate outputs against task criteria. Execute programs when a verified runtime is available. Review skill instruction and asset changes separately from execution evidence.
The three-part loop
A developer should be able to understand the product model in one scan.
Step 02
Run the system
Run a selected skill and model against the challenge cases. Trace capture can record bounded call metadata when it is enabled.
Step 03
Collect the reward signal
Evaluations score across weighted dimensions so you can see what worked, what broke, and what should change in the next iteration.
What happens after the first run
The value of the platform shows up in the second and third attempt, not just the first completion.
Inspect available metadata
When capture is enabled, inspect recorded agent, judge, tool, and sandbox events. Missing events do not prove that an action occurred.
Adjust the setup
Change prompts, model selection, decomposition, or validation where the evaluation shows weakness.
Re-run with signal
Compare scores and artifacts against earlier attempts so progress is measurable rather than intuitive.