Library

SolvNum

Plan, flight, and audit replay agree — provably.

When three teams run the same mission math on three platforms and get three different answers, the argument is never about physics — it is about arithmetic. SolvNum gives you bit-identical reals across CPU, GPU, and WASM with a cryptographic receipt, so V&V, coalition replay, and capital models stop fighting floating-point drift.

End the “works on my machine” deadlock in mission replay and digital-thread workflows.
One hash proves agreement between lab, field, and regulator replay environments.
Shrink re-validation cycles when hardware or cloud target changes.
Vertical SKUs for defence, finance, medical, and embedded — same primitive, different compliance story.

Product documentation

Multiple hardware platforms - server, embedded, GPU, browser - all hashing to the same SHA-256.

Imagine yourself in these moments. Same product, different industries.

Defence program office

Plan, flight, and replay are three different numbers.

The mission you approved on the screen, the mission the platform actually flew, and the mission the incident-review team replays in the lab are three slightly different sets of numbers — because they ran on three different computers with three different math libraries.

Imagine a single deterministic arithmetic substrate where plan = flight = replay, by construction, on any hardware. One SHA-256 receipt proves it. A cross-target equivalence argument that fits on one line of a DO-178C certification package.

Digital platform

The ML decision that flips on a different machine.

A content-moderation pipeline ingests model logits, runs calibration, applies policy rules, and emits a decision. The same logits fed into the same pipeline produce subtly different confidence scores on CPU vs GPU vs on-device — because platform math-library calls disagree in the last few ULP. At borderline cases, the decision flips.

Imagine a decision layer where every party — platform, regulator, researcher — re-derives the same decision from the same input. Zero receipt drifts across 1,000,000 decisions. EU AI Act Article 12 ready.

Industrial plant

The digital twin that cries wolf.

A fleet of 200 CNC machining centres runs a vibration model on the edge PLC and a matching twin in the cloud. When they diverge, maintenance is dispatched. Today, “diverge” usually means the two solvers disagreed on the last digits of an eigenvalue — not that a bearing is failing. The team learns to ignore alerts.

Imagine a twin where numerical divergence is zero by construction. Any divergence the maintenance system flags is a real physical change in the machine. 0% false-alarm rate from numerical mismatch. 100% fault-detection recall.