Before you dive in
What you’ll find in the original
- Bind each task's goal, measurements, exported features, contact rules, and per-layer inputs explicitly.
- Separate model-selected intent, motion family, and atomic input from deterministic simulation and execution.
- Save controls and states, then replay them without model calls to check the resulting state and task outcome.
Worth knowing
The repository provides one successful recorded example for each of three tasks, not a success-rate benchmark. Two examples used Jev through OpenRouter and one used TypeSafe directly; JevMade did not run LIBERO.