# D-JEPA: granular-relational

- Architecture: `granular`.
- Tensor count: 30.
- Includes pretrained predictor weights: false.
- Original checkpoint SHA256: `8f613280f8084e8c6403e42db138f73e01b4c43b9235e62bfd36734727836895`.
- Exported tensor-file SHA256: `a7f2b70c1d1d853cf5dbd391f3560828f28015a72092938a978dff1a0e497504`.

Load `model.pt` with `torch.load(..., map_location="cpu", weights_only=True)`. Use the matching D-JEPA code profile and JSON configuration. Tensor values are unchanged; metadata and optimizer state have been separated.

Input width: 2305. Gate: `base_minus_refined`, strict threshold `-0.06998290770476866`; quantization decimals: `None`. Candidate IDs and feature preprocessing must match the task profile.

Project license is selected by the authors before public upload.