Legal

What we claim — and how to check it

Benchmarks, product pages and documentation on this site follow the rules below. If a statement ever seems to break them, tell us — that is a bug.

Last updated: 18 July 2026 · OPS2.AI LTD, United Kingdom, Company No. 17081809, registered address 71-75 Shelton Street, Covent Garden, London, WC2H 9JQ, United Kingdom. Contact: nlsys.io@proton.me.

Claim policy

Every claim is bound to something you can check

Published numbers are executed runs

Benchmark tables consolidate release-evidence runs, not estimates. Negative gain counts as failure; an exact 0.0 dB is a declared abstention — the engine returning the input unchanged rather than pretending an improvement. Abstentions are printed in the tables, not hidden.

Models are judged in free-run

Compiled models are evaluated the way a digital twin runs — feeding themselves their own predictions plus known inputs — because one-step scores can hide error accumulation.

Refusal is a feature

On near-clean input, a do-no-harm criterion backs the filter off to the raw signal. While a channel is still calibrating, the API flags it and may return a clean passthrough rather than an invented signal. Where the data does not support a claim, the model says so.

Support boundaries are stated

Every compiled model ships with a support boundary: where it applies — and where it refuses. Its Passport includes a claim boundary — what may and may not be said about the model.

Artifacts are verifiable

A compiled artifact carries a manifest, checksums and a Passport hash. Claims travel with the artifact, and the artifact can be checked — not taken on faith.

Challenge us

Send us a case: we run it against open baselines and publish both numbers, win or lose. Send a case.

Status labels mean something

Early access — the product is live, the service layer is still being shaped with early users, and for NDC the benchmark story is still being expanded with external, hash-bound runs.

What we do not claim

We do not claim improvement on every signal — abstention exists precisely because some signals offer nothing to remove. Our benchmark tables follow our published protocol and are not commercial-competitor comparisons. Control is supervisory and advisory — we compute setpoints, your system executes; autonomous actuation is not sold. Available for civil and scientific applications only.