DeterministicML

Release: Beta — trial licensing available; compute API still in development

Bit-identical ML inference across CPU, GPU, ARM, and WASM

DeterministicML applies SolvNum-style bounded-update deterministic arithmetic to ML inference so the same model produces bit-identical outputs on every platform — x86, ARM, NVIDIA GPU, and WebAssembly.

This addresses EU AI Act reproducibility requirements, FDA 510(k) submission consistency, and any deployment where "same input, same output, every machine" is a regulatory requirement. The production wheel ships Ed25519 licensing; the inference operator set is being built on the SolvNum deterministic-arithmetic substrate.

What it does

Machine learning inference on different hardware (x86 laptop, ARM edge device, GPU server) usually produces slightly different floating-point results for the same model and input. For most apps that is fine. For regulated deployments — EU AI Act reproducibility, FDA submissions, defense certification — "close enough" is not acceptable.

DeterministicML wraps inference operators with SolvNum-style deterministic arithmetic so the same input yields bit-identical output on every platform. The licensing stack ships today; the inference operator set is expanding on the SolvNum substrate.

When to use it

Pair with SolvNum when your compliance story requires reproducible ML outputs, not just reproducible simulation math.

See Examples for runnable code.

Who it is for

• Regulated ML deployments requiring cross-platform inference reproducibility
• EU AI Act compliance programs needing deterministic model outputs
• Medical imaging and clinical decision support with audit-trail requirements

What ships

• Python wheel with licensing (deterministicml)
• C library scaffold — inference engine pending productization