Engineering & Architecture Brief Published: 2026-11-15
A constrained multi-body solver produces two kinds of numbers, and only one of them is measured. Constraint residuals, iteration counts, complementarity gaps and exit codes are values the integrator actually computed at each step. Smooth joint trajectories, energy curves and green health tiles are reconstructions built on top of them. When a part in a holonomic chain fails, these two layers are the first things to diverge, and a reviewer who reads only the reconstructed layer will sign off on a model that has already stopped satisfying its own equations.
What the solver actually logged At every step the state comes from solving the equations of motion together with the constraint equations, through Lagrange multipliers, a reduced coordinate set, or a projected iterative scheme for contact. What the solver genuinely knows at that point is short: the position-level and velocity-level constraint residuals, the multiplier values, how many iterations it spent and why it stopped. The reason it stopped matters more than any plot. A step that exits because it met tolerance is a measurement. A step that exits because it hit the iteration cap is a guess that shares the same data type, and most dashboards draw both with the same line weight and colour.
How a failed link hides itself When one link in a closed kinematic loop fractures or loses stiffness, the constraint Jacobian loses rank and the multipliers for the now redundant constraints stop being unique, but the solver keeps returning plausible joint angles because a least-squares projection always returns something, so the trajectory panel stays smooth while the reaction force panel shows the arbitrary split of load between constraints that no longer exist.
Degraded mode is something you model, not something the solver gives you A system keeps working after a part fails only if the model already contains the degraded topology and a way to switch into it. In practice that means watching the Jacobian's conditioning with a rank-revealing factorisation, detecting the loss of rank as an event rather than a slow drift, and then either removing the redundant constraint or swapping it for a compliant one with explicit stiffness and damping. Compliance keeps the integrator stable through the singular configuration, but it changes the physics: as an estimate, regularisation of this kind moves reaction forces near the singularity by a few percent, and that shift belongs in the report as a modelling assumption, not as a result. Functional safety standards such as IEC 61508 already demand that fault detection and the transition into a safe state be shown and traceable, and that requirement covers the simulation evidence too: the switch event, the residual before and after, and the formulation used afterwards all have to be logged rather than inferred from a curve that happens to look continuous.
A rule for the review Treat a failure window as measured only if every step in it exited on tolerance rather than on the iteration cap and kept the position-level constraint residual below 1 µm. Steps that fail either test are inference, and the report should say so plainly, even when the plotted values look entirely reasonable. Reaction loads at the failed joint are inference by construction, because they depend on how the redundancy was resolved, so they should be quoted alongside the regularisation that produced them. The cost shows up in real-time co-simulation, where a tighter tolerance or a refactorisation at the moment of failure can blow the frame budget. If that happens, your next decision is whether the hardware loop runs on deadline with flagged, unverified steps, or slows down until every step it shows has actually been measured.
Technical FAQ: Can energy drift on the dashboard stand in for constraint health? No. Energy is a derived quantity, and a projection that quietly violates constraints can still keep it almost flat, especially with damping or compliant contact soaking up the error. Read the logged residuals and exit codes directly, and treat the energy curve as a cross-check rather than evidence.