Frozen paper-only assets needed to reproduce the Phase 2 canid and Arabis analyses after Betty/Sesame retirement. These are not general-purpose user models; use the six separately versioned model releases for new inference.
The checkpoint files retain the exact trained state_dict and loading metadata while omitting optimizer and trainer state. The fastrho reproduce/checkpoints.json registry records both each downloadable SHA-256 and the original full training-checkpoint SHA-256. SHA256SUMS identifies every release asset.