Architecture overview
omnibias is a uv workspace monorepo built around one principle: the math has a single source of truth, and everything else specializes it.
The dependency graph
Package tiers
| Tier | Packages | Status |
|---|---|---|
| Stable workspace | core, torch, jax, ferminet | stable |
| Keras unified backend | keras | alpha |
| Beta extensions | pinn, fields, geometry | beta |
| Alpha extensions | qpinn, curvature, symbolic, score, fractional, binary, boolean, spiking, hopfield, verify, dynamics | alpha |
The Stability matrix is the source of truth for support guarantees.
The three registers
omnibias runs one derivative tower in three registers:
- Differentiable — the fast, autodiff-friendly path used in training.
- Rigorous — outward-rounded interval / affine / Taylor-model arithmetic that produces provable enclosures.
- Formal — a Mathlib-free Lean 4 kernel that re-checks finite, rational obligations.
The verified substrate and the formal bridge are pure-Python / Lean-core and never import a backend.
Design rules
These are the invariants that keep the architecture honest:
- The core is pure Python — it never imports torch, jax, tensorflow, or keras.
- Polynomial coefficients are not forked per backend. Every backend imports
them from
omnibias.core.polynomials. - Backends select the closed-form path via a class marker (
_omnibias_dispatch) rather than importing concrete field classes — which is why the foundational packages never import a downstream package. - New field-level ops are torch + jax; Keras users still get bit-identical
activation-level math through
OperatorBlock.
Where to look
| You want… | Read |
|---|---|
| The activation math | Core |
| How a backend specializes it | Backends |
| The field & operator substrate | Field substrate |
| Rigorous enclosures | Certified register |
| The Lean gate | The formal loop |