Extension Guide

driftflow extensions are explicit polymorphic collaborators. Add behavior by implementing a contract; do not mutate package-level registries or branch on class names inside the sampler.

New drift model

Implement DriftModel.evaluate(state, time) and return DriftState. Validate any external tensor-to-matrix conversion before constructing the result.

New noise schedule

Implement NoiseSchedule.value_at(time). Reject invalid configuration in the constructor and return finite, non-negative values.

New controller

Implement AdaptiveController.choose_timestep. The result must be an AdaptiveDecision with a positive timestep and finite diagnostics. Expose eps_bound and max_steps so the sampler can enforce safety limits.

New solver

Implement Solver.name and Solver.step. Solvers must preserve graph shapes, return finite values, and use the supplied random generator. Register instances on a local SolverRegistry.

New decoder

Implement GraphDecoder.decode and return an immutable adjacency matrix.

Performance backends

The public contracts intentionally separate state/control from numerical storage. A future NumPy or tensor backend can implement the same contracts as an optional package without changing the adaptive controller or sampler.