omnibias is now public. The current workspace contains 16 primitive distributions for derivative algebra, backend operators, field calculus, trainable decisions, and scoped verification.
Forward spatial derivatives, ordinary parameter gradients
Supported activation identities, Taylor jets, and direct contractions provide alternatives to nested high-order spatial autodiff. The resulting tensor expressions retain gradients into the model parameters.
Start with a runnable example
The documentation targets the current RC/alpha release family with exact version pins. Try one activation derivative, then carry a fourth spatial derivative into a training loss. A derivative demonstration is the start of a validated model, not a promise of universal PDE convergence.
Choose a complete scientific ML workflow
Try the fourth-order beam PINN, Poisson PINN, or parameter-calibration notebook. Each includes source, recorded results, and independent numerical checks.
One substrate, separate applications
Application solvers and experiments have separate repositories. Package maturity, licenses, and compatibility belong to the version you install.