Designed a Lipschitz-continuous adaptive control architecture (LCA-RLS-KAN) whose Lyapunov certificate carries no persistent-excitation requirement — and verified that the benchmark is an unexcited regime, its feature Gram holding numerical rank seven of ten. Over a 50-seed Monte-Carlo under adversarial parameter drift the controller improves on Linear Time-Varying Model Predictive Control in aggregate tracking at roughly half the per-step cost, on a branch-free evaluation with an a priori worst-case bound, while conceding a 1.6-3.6x actuator total-variation advantage to the optimizer.

Asks what a symbolic Kolmogorov-Arnold read-out actually contributes to the distillation of a Model Predictive Control law, on the Johansson quadruple-tank benchmark in both minimum- and non-minimum-phase regimes. Establishes that the symbolic extraction, rather than the network, dominates the approximation error; that imposing negative feedback as an affine inequality inside a convex read-out removes a failure mode occurring on up to 31% of the operating box by construction; and reports the negative result that the KAN-selected monomial support is statistically indistinguishable from orthogonal matching pursuit.