Pre-registered negative-results study testing whether canonical dynamical systems can be classified from observable time-series features alone. Twenty-one systems spanning seven dynamical families were embedded using 22 extracted features across spectral, temporal, recurrence-quantification, transient-response, morphology, and embedding domains.The experiment failed its pre-registered success threshold. The strongest failure occurred within the Gray-Scott reaction-diffusion family itself, where systems generated by the same governing PDE failed to cluster together under parameter variation.Results demonstrate that parameter variation within a fixed mechanism family can overwhelm observable feature similarity in low-dimensional embedding space. A companion naive-feature experiment produced complete collapse into a single dominant cluster.The paper discusses why feature-based dynamical taxonomy fails under projection loss, observable dependence, and parameter sensitivity, and outlines what types of invariant representations may be required for robust classification.
Thomas S. Mitchell (Fri,) studied this question.