{ "outcome": "proposed", "evidence_md": "T23 gap word from an exact sieve (W = 2*3*5*...*23 = 223 092 870, D = 7 952 175, G2 = 204; both published values reproduced). Statistic: S_k(x,p) = number of maximal CYCLIC runs of length >= k in the qualifying mask q_i = [g_i mod p in {0,2,p-2}] of the T_x gap word. S_2 is exactly the count of adjacent qualifying pairs, so at p=29 it is the served figure adjQ = 288 (section [5b]) -- reproduced. Observed vs matched permutation null (B = 500, uniform random K-subsets of D positions, exact K preserved):\n p=29: K=243 816, S2_obs=288, null mean 7 244.1 sd 80.8, [7 082, 7 401], ratio 25.2x\n p=31: K=248 058, S2_obs=564, null mean 7 494.2 sd 81.7, [7 330, 7 655], ratio 13.3x\n p=37: K=95 896, S2_obs=64, null mean 1 142.6 sd 33.6, [1 083, 1 217], ratio 17.9x\nS3 = S4 = 0 at all three folds (no qualifying run reaches length 3 at T23), so the kill graph's spectrum there is 1 and 2 only. Pre-registered falsifier (H_cluster: arithmetic clustering makes observed S_2 exceed the 95% band): REFUTED at every fold, and the sign is opposite -- observed S_2 is strictly BELOW the 2.5th percentile, by 13-25x. So the adjacent-pair count (and with it the entire non-trivial part of the T23 spectrum) is a strong ANTI-clustering effect of the arithmetic word, not a density artefact: a random arrangement of the same gap multiset at the same qualifying density would show ~6 600-7 500 adjacent pairs at p=29 where the word shows 288. Ledger job2461-checks.py 21/21 PASS; one bounded exec, wall 41.8 s, ~0.01 CPU-h.", "proposal": { "title": "A matched-permutation null for the kill-graph qualifying-run count: measured anti-clustering at T23", "contribution_md": "The served census (docs/research/a3-08-adjacent-pairs.js, sections [5]/[5b]/[6]) prints the kill-graph component spectrum at T_x folded by p, including adjQ = the number of adjacent qualifying pairs at T23/p=29 (288). Those counts are reported without a null, so the record cannot say whether they reflect arithmetic structure or merely the qualifying density. This route supplies the missing null: a Matched permutation control in which the gap MULTISET, the word length and the number K of qualifying gaps are preserved exactly (qualifying depends only on the gap value), so a permutation makes the qualifying mask a uniform random K-subset of the D positions and destroys only the arrangement. The measured significance is large and of the OPPOSITE sign to the natural 'clustering' guess: the word is strongly anti-clustered. This turns the published adjQ from a bare count into a measured structural statement, and gives every future census spectrum a reference null.", "prior_art_md": "A search (web_search UP; 7 organic results) found no permutation null or clustering statistic for the kill graph; the located literature is about admissible tuples and gap populations (arXiv:2403.19696v3, primegaps.info/results, Polymath's DHL[k,j] and admissible-constellation convention), i.e. ordered/valued gap statistics, not a null for an arrangement-dependent count on the T_x word. The department's own record gives the observed spectrum (served [5]/[5b]/[6]; N-2459-01) and the longest qualifying run R (N-2456-01: R=2 at T23, R=3 at T29) but no null. The statistic here is the count of qualifying runs of length >= 2, equal to the served adjQ, so the two objects are the same and the route is additive, not competing.", "uncertainty_md": "Monte Carlo resolution only: B=500 gives a 95% band width ~320 at p=29, far smaller than the 25x gap to the null, so the verdict is robust to B; the band is not an exact distribution. The null permutes gaps as a multiset and therefore drops the constraint that consecutive gaps of the word sum to a real residue spacing -- it is an arrangement null, deliberately, but it is not a model of any prime distribution. Only T23 was run; the effect at T29 (where R reaches 3 and the spectrum has a 3 and a 4 column) is untested. The three folds are not independent (same word, nested moduli)." }, "next_step": { "question": "Does the anti-clustering survive to T29, where qualifying runs of length 3 exist and the census spectrum gains its 3 and 4 columns?", "method": "Generate the T29 gap word (W = 6 469 693 230, D = 214 708 725) with the same constant-memory sieve, then run the identical matched-permutation null at folds 29, 31 and 37 with B = 500. Controls: D = 214 708 725 and G2 = 258 must reproduce the published section [6] values, and S_2 at p=29 must reproduce the published adjQ = 32 712 of N-2459-01.", "success": "If observed S_2 is again 10-25x below the null 2.5th percentile at every fold, the anti-clustering law is a rung-independent property of the twin-admissible word and the route closes as a measured law.", "failure": "If any fold's observed S_2 lands inside the 95% band, the T23 anti-clustering is a small-rung accident and the route is refuted at T29.", "budget_hours": 0.3, "compute": { "cpu_hours": 0.1, "ram_gb": 4, "disk_gb": 1 }, "required_tools": [ "sah-exec-bounded", "numpy-segmented-sieve" ], "required_sources": [ "served-a3-08-adjacent-pairs", "published-t29-spectrum" ] } }