Choose a numerical method
In Studio → Test, within the Full Car advanced 14-DOF experiment, choose original mechanics with fixed RK4, optimized mechanics with fixed RK4, or optimized mechanics with adaptive RK45. All use the same generic fixed-rack Double-Wishbone vehicle. The applied assembly is not mapped to this model. Ordinary Ride and Handling workflows keep their existing solvers.
The original mechanics and fixed RK4 remain the qualified reference. The optimized engine projects Newton–Euler inertial forces through the same physical material Jacobians, preserving all 14 mechanical degrees of freedom, full body inertia products, wheel inertias, gyroscopic terms and physical forces. Independently qualified equivalence is not a promise of bitwise identity between backends.
Adaptive accuracy
RK45 is the seven-stage Dormand–Prince embedded 5(4) pair. Each of the 28 physical state components must satisfy an absolute-plus-relative local error scale; acceptance uses the maximum normalized component. The default relative tolerance is 1e-6, with absolute position/jounce and angle/spin tolerances of 1e-10 m/rad, and linear/angular rate tolerances of 1e-8 m/s or rad/s. The tight refinement preset uses 1e-8 relative and absolute scales 100 times smaller.
The separately qualified ML-62 C++ / WebAssembly backend adds fixed RK4 and adaptive RK45 choices with Solver Analysis v3 build identity and request timing. Seenative physics, timing scope and evidence. The original default and these TypeScript v1/v2 contracts remain available.
These local controls do not bound global trajectory error. The separate qualification compares actual histories against an independent Cartesian Newton–Euler reference and a high-order extrapolated integrator. A small acceleration algebraic residual only checks the linear solve. RK45 is explicit and makes no stiff-solver guarantee.
The requested output grid is independent of adaptive internal steps. Steps end at every requested output, the pulse start/end and final time. Plots contain actual endpoint states, without output interpolation. Accepted/rejected steps, RHS evaluations and minimum/maximum accepted widths are recorded. Damping loss uses positive three-node Gauss–Legendre quadrature on endpoint Hermite states; its work/energy accuracy is independently checked, without clipping the loss.
Internal widths are bounded by configured minimum/maximum values, with default limits 1e-7 s and 0.008 s. Exact boundary clipping can produce a smaller final remainder. Limits are 4000 attempted steps, 1000 rejected attempts, 28000 RHS evaluations, 401 output samples and 300 seconds integration wall. Physical geometry, conditioning, support and finite-state failures terminate the request; they are not disguised as tolerance retries.
Compare completed runs
Run each choice explicitly. The session retains at most three completed runs and never starts an expensive comparison automatically. Physical response appears first; expand Solver Analysis comparison below it for actual wall time and evaluation charts, matching status and exports.
Speedup is shown only for identical model definition, initialization, scenario, amplitude, duration and requested times. Accuracy settings remain visible and may differ. Wall time includes dynamics, support checks, integration, sampling, positive damping quadrature and Worker yields. Compilation, equilibrium and result assembly are reported separately. Runtime, hardware load and output density affect performance; an adaptive method can be slower on a densely sampled short experiment.
Interactive runs do not execute an independent reference in the background. Reference discrepancy is therefore labelled unavailable. An inter-solver difference, including a difference from coarse RK4, is not absolute numerical error.
Ownership, cancellation and exports
Cancel terminates the request-owned Worker, including during rejection sequences. Failed, cancelled, late or stale requests cannot replace completed evidence. Changing numerical controls marks the prior record stale. Replacing the project clears experimental configuration and session comparisons; these remain excluded from workspace-project.v2.
Export individual CSV, diagnostics and provenance, or comparison CSV/JSON. Optimized records identify backend, integrator, tolerances, requested times, internal counts, runtime and reference status through separate v2 contracts. Original v1 records and their fixed-RK4 invariants are retained.
Scientific evidence and applicability
ML-61 independently recomputed numerical qualification extends the original transient evidence. Both remain distinct. The physical envelope is unchanged: small motion, four continuously loaded vertical supports, signed bounded pulses and a generic unmeasured vehicle. No nonlinear tire-slip, steering actuator, imported-road coupling or measured vehicle correlation is claimed.