Version your scenes. Replay your evaluations.

A scene record connects the objects, target arrangement and station configuration to the task. Every episode can refer back to the condition it started from.

Review the intended setup alongside the observed state. Keep the policy version, recording and result attached so the experiment remains useful after the session ends.

Create a revision from a baseline scene. Move the target, turn it, introduce clutter, substitute an object or change the lighting setup. Preserve the original condition and name the difference.

Use those revisions to build a test suite that answers a specific question. Compare matched conditions across policies, then examine combinations when you want to test interactions.

Compare scene versions together with their episodes. Inspect the setup difference, policy checkpoint and outcome, then open the evidence behind the score.

A change in the rubric or station configuration belongs in that comparison too. The record gives the team a common reference when a result improves or regresses.

Select a saved condition and re-project its placement guidance. The operator restores the objects, the station checks the scene, and the new episode joins the same evaluation history.

Use replay to investigate a failure, test a fix or compare a checkpoint against the cases your team wants to retain.

Keep important scenes in a shared test library. Assign them to stations, organize runs and review results with the people responsible for the task.

A failure should become a reusable test case. Scene history makes it possible to find the setup, understand the previous result and run the question again.

See scene history in a demo or read the guide.