Linear models are too weak. Neural nets can't explain themselves. Hand-built equations break the moment your system drifts. Upload your own data — NDC compiles it into a runtime model you can run, verify and hand off, with a Passport that states exactly what it does and doesn't know.
Pricing: part of the Analytics layer — compiles, Passports and artifacts spend plan credits; top-up packs from $149 (never expire). Full pricing →
The problem
Linear models are too weak for real nonlinear behaviour. Symbolic and hand-built models are brittle under drift, hysteresis and regime changes. Black-box fits can't explain what they learned. And even a model that works is its own project to hand to your team as something they can verify, run and trust.
Trained ML platforms answer this with a black box: accurate inside the data it was trained on, silent about where it stops being valid, and hard to hand to a reviewer as an inspectable object. NDC takes the opposite contract — a bounded model with a Passport that states its support, shipped as a verifiable artifact.
How it works
NDC compiles bounded GP-native runtime models from trajectories and ships them with a Passport, a manifest and support-aware rollout.
Ordered trajectories, optional controls/exogenous inputs, sample period and units go in. NDC normalizes on the train portion only and selects a bounded GP-native candidate family by held-out free-run.
The model is evaluated the way a digital twin runs — feeding itself its own predictions plus known inputs — because one-step scores can hide error accumulation.
You get a runtime artifact, a Nonlinearity Passport, manifest and checksums. ndc verify checks manifest, checksums, Passport hash, runtime lift and redaction boundaries. Then export or update.
Evidence
NDC v0.45.1 is a controlled commercial preview of a bounded GP-native nonlinear dynamics compiler. On the six bound civil-core official nonlinear benchmark cases, it preserves 6/6 wins versus corrected BL-02 and the adjudicated stable non-PySINDy B1 matrix under official-test free-run metrics.
| Case | NDC free-run NRMSE | Best stable non-PySINDy baseline | NDC / baseline | |
|---|---|---|---|---|
| PUB-001 | 0.0204 | 0.0547 · lightweight nonlinear state-space | 0.37× | WIN |
| PUB-002 | 0.0253 | 0.0959 · lightweight nonlinear state-space | 0.26× | WIN |
| PUB-004 | 0.2168 | 0.3174 · EDMD/DMD | 0.68× | WIN |
| PUB-005 | 0.0561 | 0.5577 · EDMD/DMD | 0.10× | WIN |
| PUB-006 | 0.2583 | 0.2838 · RLS/Kalman tracker | 0.91× | WIN |
| PUB-008 | 0.1103 | 0.1794 · EDMD/DMD | 0.61× | WIN |
Lower NRMSE is better; ratio is NDC ÷ baseline. Geomean product/B1 = 0.395. Footnote: v0.45.1 contains a direct integrated PUB-002 production artifact and hash-bound carried-forward production metrics for the other accepted benchmark routes. Real PySINDy is tracked as a non-blocking stress baseline; all-PySINDy dominance is not claimed because PUB-008 did not complete. NDC is not positioned as a replacement for every method. How we measure → Skeptical? Send us a case — we run it against open baselines, PySINDy included, and publish both numbers, win or lose.
What you get
The engine is the model; the Passport is its identity card; manifest and checksums are the seal; the rollout report is the test-drive.
| Route / basis | Selected nonlinear route and bounded basis policy. |
| Features | Channel A / B / C feature counts, raw-power policy. |
| Support / OOD | Where the model applies — and where it refuses. |
| Hashes | model_hash = manifest = sha256(model.json). |
| Claim boundary | What may and may not be said about the model. |
A compiled artifact isn't the end of the line: deployed back into a live session, it becomes a supervised twin that stays honest as conditions change — see Twin Ops.
Use cases
Cycling data → an executable degradation/dynamics model, with support and uncertainty.
Lots of nonlinearity, lags and external inputs — a strong second proof pack.
Historical sensor logs → a nonlinear digital twin without a full physical PDE model.
Drift and nonlinear actuator response → feedback-free offline what-if models.
From experiment to an executable model your team can run and inspect.
New materials, processes and rigs — a fast path from trajectories to a verified runtime.
Three ways to start
Self-service is a standing path on your own data — not a fallback while you wait for a pilot slot.
Upload JSON/CSV, compile an artifact, get the Nonlinearity Passport and a free-run report, with limited export. No pilot and no sales call to start.
Compiles spend credits from your Analytics plan; top-up packs from $149 — never expire.
Buy S pack — $149Data QA, multiple candidate runs, baseline comparisons, custom support analysis, full export and a replay session — fixed fee per dataset/project.
$4,900 fixed per dataset — 4-week scope; civil-use check first, then work starts or you get a full refund. Larger scopes — priced by quote.
Book the pilot — $4,900Private compiler deployment or a hosted private workspace, signed artifacts, an audit report and integration support — SLA on compile jobs, not on safety-critical operation.
Self-service runs on compile-credit packages (S/M/L by search budget) plus an active-artifact subscription — tens of $ per artifact/month*, with rollout/verify quotas included. * indicative starting range, validated in pilots — not a published list price. See Pricing for the full model.
Cost context: a typical MATLAB nonlinear system-identification stack — MATLAB + System Identification Toolbox + the Deep Learning / Statistics / Optimization toolboxes its nonlinear estimators require — lists at roughly $2,000–2,600 per engineer per year at published Jan-2024 list prices. A different tool with a different scope; NDC prices by artifact, not by seat.
Scope & controlled preview
NDC v0.45.1 is for civil offline identification and simulation — not a safety-critical controller. It is not for safety-critical closed-loop control, autonomous actuation, military/targeting/surveillance, RF/radar/antenna optimization, or certified medical use.
Upload your own data — no pilot, no sales call.
4-week scoped engagement, fixed fee per dataset.