Test-time rewards · collection audit

Amazon C11 distillation pilot

Inspect adaptive stopping, filtered reward panels, complete rubric-writer trajectories, and published SFT examples. Held-out reviews and oracle scores are never included in model-visible rows.

Panel coverage

Only finite, sufficiently varied reward panels enter adaptive selection and SFT publication.

Filtered panel audit

Reviewer explorer

Choose a reviewer and candidate panel.

Adaptive-stop metrics



        

Writer trajectory

Assistant reasoning

Assistant edit

Rubric after turn

Model-visible messages

Published SFT rows


      

Truncation recoveries