Do chaotic-source effects transfer across large-scale optimizer consumers? A controlled source-substitution study

Chaotic sequences are often substituted for pseudo-random numbers in metaheuristics, but source changes are commonly entangled with optimizer changes. We test whether source effects and rankings transfer across benchmark suites and consumer interfaces when non-source decisions are held fixed. Six studies at optimization decision dimension D = 1000 analyze 7430 exact-budget runs on CEC2013 and CEC2010. Studies A, B, and D use a trajectory-reuse (TR) consumer, whereas Study C uses explicitly matched two-scalar DE and PSO consumers. Study E adds canonical coordinate-wise controls, and Study F compares logistic values with an i.i.d. Beta ( 1 / 2,1 / 2 ) marginal control. Isolated random streams, paired seeds, and exact 3 × 1 0 6 -evaluation budgets preserve the intended contrasts. The TR source order is largely repeated across suites (Kendall τ b = 0.667 ), but reverses at the two-scalar DE ( τ b = − 1.000 ) and PSO ( τ b = − 0.913 ) interfaces. In the canonical controls, logistic records 2/4/2 win/loss/not-significant outcomes for DE and 0/6/2 for PSO. The matched-marginal audit leaves 13 of 16 comparisons not significant, with three mixed differences. At 50 pairs per cell, fresh uniform improves one of three targeted functions; no-cache Rössler is not significant on all three; and temporal shuffling improves Rössler once and pendulum twice. The evidence is consistent with source–interface dependence within the tested two-scalar and canonical coordinate-wise consumer interfaces, not a portable benefit of nominal source dimension, physical provenance, or chaotic origin.