# Preregistered matched gap-order experiment Before any new permutation run. Actual cubic-window R values were already observed in our return233; this is a new, explicitly conditional comparison, not a blind discovery of those values. Question: does actual arithmetic gap order reduce the Charlier zero-detector energy beyond its exact gap multiset? Primary: T23,H12168,m8. Secondary T13,H2202,m6 andT19,H6864,m7. Statistic R=E_a K_m(S(a))²/E_m(mu), with S periodic interval survivor count, mu=HD/W and the exact polynomial family of233/234. Lower is the proposed favorable direction. Null:127 independently seeded Fisher-Yates permutations of the full cyclic gap multiset at each level. Preserve D,W,all gap frequencies,mean gap and single-gap maximum exactly. Seeds202609130000+1000*z+rep,rep0..126;mt19937_64 draws with rejection to unbiased bounded integers. Counts/histograms exact integers, polynomial expectations exact rational numbers. Primary decision: advance an arithmetic-order improvement only if lower-tail Monte Carlo p=(1+#null R<=actual R)/128<=.01 AND R_actual/median_null<=.90. Otherwise do not advance that claim. Report secondary levels without pooling or reselecting the primary. Resolution1/128; a10percent effect is the minimum relevant magnitude, not a power guarantee. This conditional randomization test evaluates exchangeability relative to the specified gap-order null, not a stochastic law for arithmetic. Checks: recover exact all-translate histograms from233 for each actual tile; every shuffled gap multiset identical; small cyclic histogram from independent direct sliding counts; a separately seeded null pseudo-observation scored against the127reference draws at each level as a calibration diagnostic (no claim from three samples). No asymptotic bound or direction filing solely from a hit. Cost budget: one CPU,2GiB address cap,300CPU seconds per engine run,<=10GB disk; session deadline takes priority. If calibration/reproduction fails, invalidate the comparison rather than change the gate.