Results

Every number here came from a run

We publish results together with the protocol, the dataset and the command that reproduces them. Not because we need checking up on, but because in engineering a number without a protocol is worth nothing — and we would rather hand you something usable straight away.

See the results Download an artifact

Benchmarks

What the engine did on public data

+8.2 dB

on multi-tone vibration at 20 dB SNR — where the best classical filter goes to −9.3 dB.

6/6

free-run wins for compiled models against classical methods on public cases.

32/32

real-waveform cases carried to the end, including the ones where the engine declined and printed it.

numbers on this site without an executed run behind them.

These are engine-level results on signal benchmarks. The battery pack is a separate table with its own dataset and its own baselines, and it will be published the same way: dataset, protocol, both outcomes.

Artifact

A finished result you can download and take apart

A complete run on a public dataset: the source, the cleaned output, the change table, the dossier in machine-readable and human-readable form, the manifest and the checksums. Only the training payload is left out — and that omission is itself a file you can inspect.

Two commands check the whole package in about half a minute: one verifies the checksums, the other verifies that the report agrees with the change table. Anyone you hand the artifact to can do the same — a colleague, a customer or a reviewer.

What is in the dossier

$ sha256sum -c checksums.sha256
  cleaned.csv               OK
  changes.parquet           OK
  experiment-dossier.json   OK

$ nlsys-battery verify .
  rows in / out   421084 / 421084   ok
  row order preserved              ok
  table vs report  59754 / 59754   ok
  9 pass · 2 warn · 1 insufficient

How we word our claims

Three rules the whole site is written by

Every number is a run

No “up to”, no projections, no figures carried over from a neighbouring product. If the run has not happened, the cell stays empty.

A refusal is a result too

An exact 0.0 dB means the engine left the signal alone because the noise could not be shown to be separable. Those cases go into the table alongside the wins.

Every conclusion has a boundary

A claim always comes with the conditions under which it holds. Outside them the model declines rather than extrapolating.

Comparison

Where we are strong, and where to pick something else

	Your own script	An enterprise platform	nlsys
Cost of entry	Your own time	A sales call first	A minute to sign up, starter volume free
Scale	One person, one dataset	A project with a budget	From a single file to a campaign
Formats	Whatever you write	Broad, productised	CSV and TSV through a mapping step
Record of changes	Whatever you log	At platform level	Per cell, with reason codes
Behaviour when unsure	However you code it	Usually hidden	The refusal is printed in the report
What you keep if you leave	Your code	As a rule, nothing	Plain files and a free verifier

What we are good at

A result that survives review: you can show not only the outcome but what was done to the data.
dQ/dV peak integrity is measured and printed, not asserted.
You can start today — a minute to sign up, no procurement conversation.
Artifacts stay readable after a subscription ends: plain files plus a free verifier.

What we do not do

We do not read binary .ndax, .mpr and .res — we work with CSV and TSV. How to export.
We do not replace a LIMS, and we do not do supplier scorecards, sampling plans or lab scheduling.
We do not issue the regulatory EU Battery Passport under Regulation 2023/1542: our dossier is an engineering record of an experiment.
We do not sell autonomous actuation: Control computes setpoints, your system executes them.