Failure Categories

Category	Evidence	Impact	Next Steps
Training Maturity	exact550, official e1250/e1500/e1750 have not completed three batches	Cannot prove that the 6000 epoch goal has been achieved	Reserve a full computing window, save segments for evaluation
Data Distribution	Among 550 samples, rescue=257; clean50 performs differently from exact550	Possible deviation in profile / pose / joint coverage	Conduct single-variable paired ablation on stock / retry / rescue
Time Alignment	next-state max error=0 for processed audit, but loader modifications will change the target	Misalignment will distort both loss and closed loop	Fix the official semantics, conduct separate alignment experiments
Evaluation Topology	Both ordinary runner and trace runner yield different results for the same weight	Small sample scores cannot be directly compared	Finally use the persistent official runner
Public/Private Gap	Local control: fresh3=2/3; online T4=25.3/100	Public 3 seed does not represent private 100 seed	Add fixed, unseen seed panels
Resource Availability	Remote 4090 is shut down due to debt; PRO 6000 has disk/GPU ownership constraints	Training and synchronization windows are compressed	Persistent disk, balance alerts, recoverable checkpoints
Scaling Process	403 error for wrong account; 403 error after the second version deadline	Only one valid online T4 exists	Check unlock status and save queue id before submission
Release Boundaries	Private logs contain paths, processes, credentials, and large files	Direct copying would reveal secrets or cause oversized repositories	Only publish masked summaries and reproduction scripts