Executable model
Canonical dynamic model for simulation and rollout.
NDC Model Compiler
Provide time-ordered data, one target and the inputs available at runtime. NDC searches alternative dynamic structures, challenges them in held-out free-run and packages the selected result as an executable, verifiable artifact.
From €390 · one target · standard processing allowance · additional usage, when required, is shown in the account before launch.
Chronological observations, one target, measured inputs or controls, and stable sample timing or timestamps.
Alternative dynamics, memory and cross-signal structures are tested under the declared data boundary.
Runtime, validation evidence, Nonlinearity Passport, support boundary, manifest and checksums.
Why NDC
One-step fit can look strong while a model drifts or becomes unstable as soon as it feeds back its own output. NDC therefore records both one-step evidence and a chronological held-out free-run.
The model receives allowed future inputs, but not the true future target. Each predicted target becomes part of the next state. Accumulated error is visible rather than hidden by teacher forcing.
Illustration of a free-run comparison, not a measured benchmark. Published measurements are linked below.
Published model results
The six published cases report free-run NRMSE ratios from 0.10× to 0.91× against the declared baselines. They are recorded benchmark cases, not a guarantee for a new dataset or a statement that every case was recompiled in the present release.
Automated compile
Chronology, target availability, selected signals and usable sample count.
Align selected views and create a reproducible fit and validation boundary.
Evaluate dynamics, lags, memory, self-effects and cross-signal interactions.
Estimate coefficients, numerical rank and weakly identifiable directions.
Measure one-step and held-out free-run behaviour with accumulated error.
Package the selected runtime, Passport, support boundary and verifier.
What you receive
Canonical dynamic model for simulation and rollout.
Run the delivered model without rebuilding the compiler pipeline.
One-step and free-run metrics, baseline comparison and rollout behaviour.
Structure, nonlinear order profile, rank and uncertainty where estimable.
Observed operating ranges and explicit behaviour outside them.
Versioned request, input identity and hash-bound delivery files.
From model to action
After a model passes its declared free-run and support checks, the next product route is control: real-time SDRE for a strict update cadence or full-model NMPC where a slower optimisation loop is acceptable.
Controller compatibility still depends on measured state, actual applied commands, update cadence, declared limits and the requested objective. Unsupported configurations must remain rejected.
Pricing
NDC is a productised compile, not a consulting engagement. The account confirms the current input limits and any additional processing before the job is launched.
One model for one target under the standard processing allowance.
Questions
Yes when the file represents ordered trajectories with a target and the inputs needed for future rollout. General cleaning and gap recovery are separate automated products; battery cycler data uses the Battery workflow.
The measured response the model must reproduce—for example temperature, voltage, position, flow or a quality indicator. Each target produces a separate verified model.
No. It reports predictive relationships—signals and lags that improve prediction. Causal conclusions require a separate experimental design.
The current workspace offers “Refuse unsupported inputs” and “Refuse extrapolation”. The selected policy is included in the compile request. Follow the policy recorded in the delivered artifact; an output outside validated conditions is not a validated prediction.
NDC