Update five XL maze models to validation-selected epochs (2026-11-03)
The existing maze exports predate the XL campaign. This update adds 40 original checkpoints for five mazes, four families, and seeds 0 and 1, and points the Enoch model card at them.
Epoch selection uses only frozen validation cases, averages both seeds, and breaks ties toward the earliest epoch. TD-MPC2 averages all three standard modes and uses one selected epoch across modes. DINO/LEWM/TS use latent-planning success. The two incomplete TS validation stages were resumed before selection.
| Maze | DINO | LEWM | TS | TD-MPC2 |
|---|---|---|---|---|
| Arena | 5 | 15 | 5 | 45 |
| Maze | 25 | 5 | 10 | 35 |
| Bottleneck | 15 | 35 | 10 | 20 |
| Circle | 5 | 10 | 5 | 10 |
| Spiral | 20 | 30 | 15 | 10 |
The export includes checkpoint hashes, validation evidence, source identities, action normalization, and TD-MPC2 configs. All original checkpoint bytes and selected evaluation evidence were verified on Slurm compute nodes. Older exports and other contributors' files are preserved.