Current step——
Mean reward, last step——
Step time——
Time remaining——
Reward over training
Mean reward per step, all environments
one point per training step; infra failures excludedModel health
how far each update moves the policy
Rollout mechanics
context compaction, carried-over episodes, loop judging