Ranking versus Absolute Accuracy in Matter–Light Simulation Cost Prediction

Ordering quantum-dynamical simulations by difficulty is not equivalent to pre-dicting their absolute computational requirements. We examine this distinction using frozen matter–light product-formula benchmarks and descriptors of accessible support, state-space geometry, finite-time Krylov evolution, control-algebra dimension, analytic sufficient steps, and empirical symbolic predictions. Effective Krylov dimension has a median error-stratified Spearman association of 0.8158 with measured steps, but the descriptor comparisons use different eligible populations and cannot establish condi-tional superiority. On its strict dimension holdout, a frozen symbolic predictor retains rank correlation 0.9363 while incurring mean absolute error 626.5785 steps. Analytic sufficient-to-measured step ratios have medians 46.6087 and 7.6583 for the respective formula orders; these measure conservative looseness, not point-prediction error. The benchmark illustrates a domain-limited separation of ranking, absolute accuracy, and conservative error control. Censoring, approximate Krylov bases, and unavailable cir-cuit domains preclude a unified or hardware-level complexity interpretation.