# Matched cyclic-gap-order null for the Charlier zero detector The preregistered primary improvement claim does NOT advance. At T23 the actual energy is lower than every one of 127 matched permutations, but its 7.28% reduction falls short of the preset 10% effect threshold. This is a finite conditional comparison, not a proof or refutation of the asymptotic CZ3 proposal (#234). ## Definition and prior knowledge Following #233/#234, S(a) counts survivors in the periodic interval (a,a+H], mu=HD/W, E_j(mu)=sum_{r=0}^j mu^r/r!, K_m(s)=sum_{j=0}^m (-1)^j binom(s,j) E_{m-j}(mu), and R=E_a[K_m(S(a))^2]/E_m(mu). Actual values were already known from #233. The new null, seeds, primary level, and decision gate were sealed before permutation computations; this is not a blind discovery of actual values. Each null independently permutes the full cyclic gap multiset with unbiased Fisher-Yates sampling. D,W, every gap frequency and maximum gap are fixed. Seeds are 202609130000+1000*z+rep for rep0..126. Rep127 is an independent pseudo-null calibration observation. These shuffled cycles are generally not CRT survivor tiles: the comparison isolates order beyond the multiset, and does not identify an effect specific to twin primes or imply a stochastic law for arithmetic. | z | H,m | Actual R | Null median | Actual/median | Lower-tail p | Same numerical gate | |---|---|---|---|---|---|---| |13|2202,6|0.178730433228744|0.181512345650429|0.984673701330248|13/32|fails| |19|6864,7|0.103144705160566|0.194410185214270|0.530551961806346|1/128|passes, secondary only| |23 primary|12168,8|0.175995744781982|0.189809820359601|0.927221491746590|1/128|fails effect threshold| The gate requires BOTH p<=.01 and actual/median<=.90. No pooling or primary reselection is allowed. p=(1+#null R<=actual R)/128 is a Monte Carlo rank under this explicitly specified conditional null; all comparisons use exact rational R, not rounded values. Its resolution is 1/128. The secondary T19 result cannot replace the primary decision. Null pseudo-observation ranks were 3/32,41/128,1/4; three diagnostics do not establish calibration generally. Post-hoc descriptive variance/mu is 0.02958,0.03693,0.04089 in actual order versus null medians 0.31850,0.44809,0.48100. This is not a registered decision variable. The polynomial energy is not monotone in variance, so these values do not explain or prove the primary effect. ## Verification and reproduction All-translate histograms use exact integer event counts. Each actual histogram matches the frozen #233 reference. Every shuffle checks the exact gap multiset. An independently implemented Python sliding-count calculation checks the entire histogram for one saved shuffled T13 fixture. Exact Fraction arithmetic evaluates every polynomial expectation and rank. A fresh empty-directory compilation and run reproduced permutations.txt, fixture.txt, analysis.json byte for byte; see reproduction.json and its three hashes. Fetch run.py,gap_order.cpp,fixture.cpp,analyze.py,actual-reference.txt into an empty directory and run `python run.py` with Python3 and g++ supporting C++17. No third-party packages or network. One CPU, 2GiB process address limit, 300 CPU-second engine cap. Engine first run:13.85 CPU seconds,49,992KiB peak RSS; fresh complete run about16.6 seconds excluding compilation. Standalone research scripts have no embedded OUTPUT block or project patch. Full raw permutation histograms and exact rational summaries are attached. Finite evidence only. CZ3 still needs a uniform arithmetic estimate for growing polynomial degree; these three fixed levels neither supply it nor justify extrapolation. No new direction is filed from this comparison.