Integrate a policy¶

A new policy must obey the task's observation and action contract. Keep the simulator criterion fixed while implementing the adapter.

For the current open core, use the frozen protocol and reference trainer. Its ACT/DP/BC implementations are already integrated; a different architecture requires an explicit loader/rollout adapter. Record a new method identifier and its source changes. Do not overwrite a reference checkpoint or claim that modified runtime code passes the existing frozen manifest. Preserve task physics, observations, reset cohorts and success criteria when presenting a comparison; disclose any deliberate change to them.

Adapter boundary¶

Inspect an existing loader in src/wasman/learning/, its training script and the corresponding rollout backend. Declare camera ordering, image preprocessing, measured-state fields, output normalization, history length, action horizon, executed chunk length and reset behavior. Preserve checkpoint/config identity.

On recorded expert observations, compare model predictions, decoded targets and commands actually passed to the controller. Locate the first disagreement before starting a long closed-loop experiment. A newly introduced limiter is a controller change and must not be silently folded into an architecture comparison.

Deployment checks¶

Reload the selected checkpoint in a clean process. Verify finite predictions, chunk timing, reset of history buffers and the absence of privileged expert inputs or replacement actions. Run a bounded development rollout, retain its trace and inspect the first physical deviation. Matching offline actions does not by itself establish closed-loop success.

The open-core reference protocol uses the fixed final-budget checkpoint; an alternative selection rule defines a different experimental setting and must be reported explicitly. Report one training run as one run; reset confidence intervals do not measure training-seed variability. Record unsuccessful outcomes as well as successful demonstrations. The current revision schedules three independent trainings per reference family and task. VLA integration and additional policy families are not completed evaluations.