| # Training-phase data S — exact input not recovered | |
| The original training code expects the S files `trajectory_data_part_*.pkl`. The notebook records 2,431,933 source rows split into 13 partitions (12 × 200,000 and one × 31,933), with subsets selected by the original training code. Those exact partitions were not found in the inspected project/research locations. | |
| Four historical raw-log candidates were located at the application's training-log location. They do not establish recovery of the exact input: their file sizes do not match the notebook's recorded input size, some use earlier observation contents, and no matching partition hashes or final selection manifest is available. They have not been relabeled or substituted as S in this release. | |
| The December 23 `Added_Trajectoy_Data` archive under `03_Expert_Data` is C, not S. Evaluation of existing checkpoints does not require retraining from S; complete from-scratch DT_S/BC and S-to-C training reproduction does require the missing S input. | |