# Job #2461 — a matched-permutation null for the kill graph: the qualifying run count is strongly ANTI-clustered at T23 **Assignment.** explore / discovery / stage discover, **routeless**, lane adversarial, general mode, 1 of 1 (session `59ad3d0fd45441dd49103e37`, attempt `80a2689de7e3a9a26f17b1eef8c72987`, job **2461**). **One-line result.** The census's adjacent qualifying-pair count at T23/p=29 (the served `[5b]` figure **288**) sits **25× below** a matched permutation null (mean **7 244**, band `[7 082, 7 401]`), and the same 13–25× anti-clustering holds at p=31 and p=37 — a pre-registered "clustering" hypothesis is **refuted**, and the measured effect has the opposite sign. ## The statistic and the decision it informs The kill graph is built from the T_x gap word by `sigma' - sigma = g_i (mod p)`, `sigma ∈ {0,-2}`; a gap *qualifies* iff `g mod p ∈ {0, 2, p-2}`. Define, for the cyclic word, **S_k(x,p) = number of maximal cyclic runs of length ≥ k in the qualifying mask.** A qualifying run of length ≥ 2 is exactly the condition for a non-trivial (length ≥ 2) component, so **S_2 = the number of adjacent qualifying pairs = the served `adjQ`**, and S_3/S_4 are the spectrum tail. The census prints these counts at T5…T23, T29 (`[5]`/`[5b]`/`[6]`), but **no null**, so the record cannot decide whether they reflect arithmetic structure or just the qualifying density. That is the decision this statistic informs. ## Matched control Qualifying depends only on the gap **value**, so a uniform random permutation of the gap word preserves the word length and the gap **multiset** exactly, and therefore preserves the number **K** of qualifying gaps exactly; the qualifying mask becomes a uniform random **K-subset** of the D positions, and only the **arrangement** is destroyed. Sanity controls: the identity permutation reproduces the observed S_k, and D = **7 952 175**, G2 = **204** reproduce the published T23 values. ## Pre-registered falsifier (written before any run) *H_cluster*: the arithmetic word clusters qualifying gaps, so observed S_2 exceeds a central 95% band of B = 500 permutations. **Falsifier:** observed S_2 inside `[q₂.₅, q₉₇.₅]` ⇒ H_cluster refuted. ## Result (`job2461-cluster.log`, ledger `job2461-checks.py` **21/21 PASS**) | fold p | K (qualifying gaps) | S2 observed | S3 | S4 | null mean | null 95% band | ratio | |---|---|---|---|---|---|---|---| | 29 | 243 816 | **288** | 0 | 0 | 7 244.1 | [7 082, 7 401] | **25.2×** | | 31 | 248 058 | **564** | 0 | 0 | 7 494.2 | [7 330, 7 655] | **13.3×** | | 37 | 95 896 | **64** | 0 | 0 | 1 142.6 | [1 083, 1 217] | **17.9×** | S2(p=29) = **288** reproduces the served `[5b]` adjacent-pair figure — the statistic and the published object coincide. At **every** fold the observed value is strictly **below** the 2.5th percentile, so H_cluster is **refuted** and the measured sign is **anti-clustering**: 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. S3 = S4 = 0 everywhere, consistent with N-2456-01's R = 2 at T23: the T23 spectrum is `{1,2}` only. ## Rung - `D = 7 952 175`, `G2 = 204`, and **S2(p=29) = 288 = served `adjQ`**: **verified** (exact reproduction). - The permutation-null separation and its sign: **measured** (Monte Carlo B = 500, exact-K subset sampling). ## Gap that remains Only T23 was run. At T29 the qualifying run reaches length 3 (R = 3, N-2456-01) and the spectrum gains its 3 and 4 columns, so the anti-clustering is untested where the tail is non-empty. The null is an *arrangement* null and deliberately drops the congruence structure of consecutive gaps; it is not a model of prime distribution. The three folds are nested (same word), not independent. ## Cheapest next experiment (pre-registered, in `job2461-research.json`) Run the **identical** null on the **T29** word (W = 6 469 693 230, D = 214 708 725) at folds 29, 31, 37, with controls D = 214 708 725, G2 = 258 (served `[6]`) and S2(p=29) = served `adjQ` = **32 712** (N-2459-01). Success ⇒ rung-independent anti-clustering law; failure at any fold ⇒ refuted at T29. Cost **0.3 h / 0.1 CPU-h / 4 GB** (the T29 sieve is ~11 s; a 500-permutation sweep over D = 214 M is ~2–3 min). ## Also on the record - Prior-art search (`web_search` UP, 7 organic results) found **no** permutation null or clustering statistic for this object; the located literature is admissible-tuple/gap-population work (arXiv:2403.19696v3, primegaps.info, Polymath `DHL[k,j]`). No published number is contradicted. - 42 of @Benjaminsen's returns wait for a verdict (12 on deepseek-v4-flash), the oldest since 2026-09-11. - Usage for this attempt: the harness exposes no attributable token counts → **unmeasured/pending**. ## Files `job2461-cluster.py` (experiment, question in the header), `job2461-cluster.log` (raw output), `job2461-checks.py` (ledger), `job2461-research.json` (proposed route), `REPORT.md` (this file).