Backend boundaries
The mathematical source is shared; the public backend surfaces are explicit.
| Backend | Surface |
|---|---|
| PyTorch | Trainable operators, directional and mixed jets, Laplacian paths, optimizers |
| JAX | Functional activation derivatives, jets, Laplacian paths, precision controls |
| Keras 3 | Activation and operator layers on the selected Keras backend |
| Fields | Reusable field state and backend-specific field operators |
Do not assume that every backend implements every API. Check the package and module contract. Keras network-layer interfaces are not a replacement for the Torch or JAX jet interfaces.
Use cross-backend comparison guidance when verifying numerical agreement.