cwam

CWAM S1 action policy on FR3 clean149 (149 episodes, episode 135 removed), trained from nvidia/Cosmos3-Nano-Policy-DROID with frozen Wan2.2 latent cache (metadata-only tokenizer).

• Experiment: FR3-DROID-S1-H260-8G-P9-S42-20260919-R1
• Steps: 100,000 (checkpoints every 10k on OSS)
• Recipe: action_policy_fr3_droid_s1

Exported from DCP iter_000100000 via cosmos_framework.scripts.export_model.

See fr3_policy_export.json for dataset/cache identity receipt.