EVIDENCE & BENCHMARK
Judged by independent ground truth, not by model fit
Benchmark in development. Our public known-answer benchmark is not yet approved for publication — so instead of simulated results, this page sets out the evidence standard every published result must meet.
THE EVIDENCE STANDARD
What every published benchmark must show
No result is published without documented ground truth, metric definitions, scope, version and limitations.
The state and why it is hard to measure
The physical state being estimated, the machine system class, and why direct measurement is difficult or costly.
Inputs used — and not used
Exactly which measured signals and public parameters were used, and which data were deliberately not used, including any avoided proprietary shortcut.
Calibration kept apart from validation
Data used to calibrate the module is kept separate from the independent data used to judge it.
Estimate versus actual
An appropriate time series and error distribution, with MAE, bias and tail-error measures where applicable. Uncertainty is shown, never hidden.
Operating Domain and refusal cases
Where performance holds, where it changes, and outside-domain cases where the VSM is allowed to say it does not know.
Version, provenance and limitations
The VSM and Validation Record version, the source of every parameter, and the known limitations of the result.
BENCHMARK STATUS
Benchmark in development
We would rather show no chart than a simulated one.
Publication criteria
The first public benchmark will be published only when its ground truth, metric definitions, scope, version and limitations are documented and approved.
Validation Record preview
Target state · inputs used and not used · parameter sources · calibration data · independent validation data · error metrics · uncertainty · Operating Domain · outside-domain behaviour · version history.
When evidence is confidential
If a customer result cannot be made public, we describe the method and its boundaries — never an unsupported number or an anonymised claim we cannot substantiate.
PROPOSE A KNOWN-ANSWER CASE
Have a state you can already measure?
If you have an independent reference for a hidden load, force or state, it can become a known-answer case — the strongest test of whether existing signals can answer the question.