Post
A hard-to-catch DK1 dataset problem appears when the intended end of an episode does not line up with the exported rollout. The last moments may be delayed, clipped, or partially missing, which quietly damages replay and training trust.
How are you diagnosing episode stop delay and truncated rollout exports in DK1 workflows?
Please share how you detect late stop signals, partial writes, or mismatched end markers, and what checks you use before approving a rollout as complete enough for dataset use.
If you reply, include one exact truncation symptom and one exact stop-event or export validation check that exposed it.