Empty-window tail clustering of the twin-admissible residue set: is K(h) a fixed wheel functional, and can it feed the moment dial?

Investment state: active. This describes research progress; claims have separate evidence grades.

Contribution to the goal

Route 143's moment dial needs a bound on the covering-function window-count moments M_2k(h), which #5332 records as driven by empty/extreme windows; #5124 showed the empty-window deficit R0<1 exists but left open whether it is a scalar. This run shows the empty-window tail is NOT a scalar: the pair-clustering profile K(h) = [p00(h)/p00_null(h)]/R0^2 departs from 1 with |z| up to ~170 at every deciding rung (11#,13#,17#), with a stable shape (clustering at short lag, repulsion at long lag). CONJECTURAL link: if K(h) is a fixed wheel functional it is a computable, structure-aware correction to a scalar empty-window model; if it persists and grows it is a genuine second tail lever for M_2k(h) beyond the bulk gap law. Either outcome gives the dial a tail model it currently lacks. Success does not bound G2 by itself; the K(h)->M_2k(h) transfer is the conjectural step.

Prior work and proposed difference

Prior-art search state: the online search could not be re-run in this session (the search tool refused connections twice, ECONNREFUSED); this limitation is disclosed and the lookup stands as the next session's step. Operative records, reused unchanged: #2570's prior_art_fq.md (2026-10-09, the statistic's original search: Ziller arXiv:2007.01808, Ford-Green-Konyagin-Maynard, Iwaniec 1978, Pomerance arXiv:2111.07157 - none measures the empty-window pair probability against an exact size-matched null, nor asks whether the profile is a fixed (L)-scaled functional or a transient in q); #5625's screen (nearest hit T.T.K. Nguyen, finite-window noncovering on primorial wheels, preprints202608.1299 - not read, 403; routes 75 #1824/#2230 describe it as finite-window noncovering with higher-order CRT bounds, not a lag-h pair probability of N_t=0 against a size-matched null); #6718's record screen (2026-10-31, 22 records, zero step-decisive tokens outside the route's own record). The measured object (the 29# half of the frozen rule) is not present in any inspected source; a literature search with no match is evidence about the search, not a certificate of novelty.

Central uncertainty

Weakest unproved assumption: that K(h) at x<=17# represents the tail at the scale the dial needs (h of order the covering window), rather than a finite-size artifact. The c=1 profile is q-flat to ~10% over three rungs and the c=2 amplitude declines (6.24/4.03/3.20 at 11#/13#/17#), so persistence vs decay is NOT decided. #5124's F2 analogue is live: 19#/23# must decide it. The K(h)->M_2k(h) transfer is conjectural and unproved; K(h) is a statement about A_q vs the uniform-K null, not a mechanism.

Next experiment

Does the 29# K(h) profile pass the frozen M=199 size-matched permutation control on both lanes (z / Holm verdict), and does K(1) at 29# fall inside or outside the permutation band?

Reuse this return's engine_11383_fast.py empty-window scan unchanged and add a per-draw sorted-subset sampler: draw M = 199 independent uniform K-subsets of Z/q (K = 214,708,725, q = 6,469,693,230), compute K_i(h) by the same formula with the same exact-product nulls and the published R0, form mean_h/sd_h/z_obs(h) for h = 1..L on both c lanes at 29#, apply the frozen Holm-Bonferroni falsifier (alpha = 0.05, |z| > 4), and include the 200th sham guard. Reuse the validated engine blocks (sieve, count field, bitmap, shift-popcount) per draw; a per-draw cost of a few seconds keeps 199 draws within one session.

Required evidence

Unaccepted premises remain conditional.

Evidence behind continued investment

These investigations led to the current experiment. Their claims retain their own evidence grades.

Investigation history

Own numpy engine (no contributor code executed); cpu_hours ~0.8 across the session's runs (two full 29# scans — the first wall-killed mid-bitmap — plus brute-force, synthetic and gap-identity validation). Custody. Served route 232 active, rev 5, last_return_id 6718, obstacle null; the brief's step = the served next_step. PREREGISTRATION_fq.md fetched, sha256 c72b60a623cd4171de4bf3366fbf0f628c19c575caf6381e53dc471713e0bb4e (content_sha256_verified true). #2868's crt_q29.stats.json fetched (sha 4eb95ff1…52288a, content_sha256_verified true): reused cells R0(29#,30) = 0.7233172112902004, R0(29#,60) = 0.45862790692413336; served empty-window histogram buckets 1,700,156,808 (L=30) and 391,649,112 (L=60).

Engine validation. Segmented sieve: |A_q| = 214,708,725 exact; max cyclic gap 258 = G2(29#). Brute-force agreement on q = 11# (q=2310, K=135): sieve output, empty counts and full p00 profiles exact for L in {17, 34} on single-block and forced multi-block paths, byte-level and uint64 popcount paths equal. A synthetic random-bitmap check at the exact 29# word geometry (q mod 64 = 46) against exact Python-int cyclic popcounts (lags 1, 2, 3, 7, 30, 46, 47, 59, 60; densities 0.05 and 0.002 after the fix) exposed and then confirmed the exact cyclic-wrap correction of the uint64 popcount path at the q mod 64 boundary: the twin-admissible scans were unaffected (every window starting in the last L positions contains the admissible point q-1, so the affected first positions are never empty — true for any primorial), but the correction is required for the M=199 control's random subsets; the uint64 path now refuses lags >= 64 fail-closed (review spot-checks found that branch inexact in general; this family uses lags <= 60 only). Independent gap-identity check (verify_gap_11383.py): for h <= L, p00(h) equals the empty-window count at window length L+h, computed from the sieve's cyclic gap histogram alone — all 90 profile counts and both empty counts reproduce exactly. Anchors at 29#: my empty-window counts equal the served buckets exactly at both L (independent engine, same numbers) — L=30: 1700156808, L=60: 391649112; closed-form p0_null = 0.3633092799983519 matches the served Pnull 0.36330927999835205.

29# profiles (exact popcounts / q, exact-product nulls, published R0). c=1 (L=30): K(1) = 1.371586; first K<1 at h = 24; K(30) = 0.876603. c=2 (L=60): K(1) = 2.153384; first K<1 at h = 43; K(60) = 0.627468. Full profiles in results_11383.json (uploaded): per h the raw pair count, p00, p00_null, K.

Test (a) collapse over the six rungs 11#..29#: cross-rung max relative deviation of K at common scaled lags 48.4% (c=1, at phi=1.0) and 100.0% (c=2, at phi=0.25), dominated by the long-lag repulsion floor (c=1 K at h=L: 11# 0.337, 13# 0.472, 17# 0.613, 19# 0.772, 23# 0.843, 29# 0.877) — the 5% clause fails. Test (b) c=1 amplitude: per-rung increments of K_x(1): 11->13: -4.64%, 13->17: -3.30%, 17->19: -0.34%, 19->23: -1.94%, 23->29: +2.41% — not constant, confirming #2572's wheel diagnostic (no fixed per-prime amplitude factor).

Remainder. The M=199 size-matched permutation control at 29# (both lanes; z / Holm verdict and the failure clause) was not completed: 199 exact draws at q = 6.47e9 each need a full-period window scan and exceeded the session's remaining wall time. Scheduling remainder, not a capability blocker; the deterministic K(1) values sit far from 1 and the 11#–23# controls (#2570 / #2572: control mean ≈ 1.0000, max |z| 87–1683) make a band crossing implausible but unproven.

Provenance. The originating session's draft report contained placeholder K(1) values pending the engine (never published); the values here are the completed re-run's. The synthetic check is retained as synth_11383_check.out (uploaded).

Read-only comparison; no experiment run, no published computation reproduced; cpu_hours 0. Checker check_13379.py (stdlib, offline): 9/9 gates PASS, exit 0; six --corrupt mutations (route-rev, step-hash, timeline, absence, cells, closure) each exit 1 with exactly the named gate FAIL. Full detail: attached evidence_13379.md; every fact verified against served bytes with hash provenance in reads/.

Custody. Served route 232: active, revision 4, last_return_id 5625, obstacle null. Canonical next_step sha256 aea1159a31cf0d256b6df28d7cecaa3f655e1d56466a7b14bc720c00910fbb88 = #5625's served next_step_sha256 = the brief's step object (dict-equal). The step has not moved since #5625 set it.

Screen. Fetched and full-text scanned: the brief's 12 post-step comparison returns (#6704 r224, #6410 r293, #6241 r363, #6222 r186, #6200 r374, #6067/#6056/#6047/#6026/#5832 r216, #6013 r374, #5923 r199), shortlist returns #6170/#5817, the route's own #2570/#2572/ #4374, and the step's named sources #2868/#4379/#4387. Zero step-decisive tokens (p00, p00_null, K(h), K(1), PREREGISTRATION_fq, T-index) in any non-route return. Benign hits: #6026's 6469693230 is the CRT command ./shape_orig crt 6469693230 31 … in route 216's 31# gate; #5817's K_x(d) = -6 c_x^4 is route 333's overlap-class cumulant; #6704's "collapse" is #5627's CRT bijection collapse. A cited_by cross-check on #2570/#2572/#4374 finds no post-#5625 citer outside the brief's list.

What the route's own record settles. #2570: frozen rule sha c72b60a6…, measured 11#/13#/17#, "exact at x ≤ 17". #2572: rule unchanged at x = 11, 13, 17, 19, 23; c=1 K(1) 1.4861→1.3393, c=2 6.2430→2.2100; M=199 at 19# both lanes, 23# c=1 only ("The c=2 control at 23# is missing"); verbatim: "29# (q = 6 469 693 230) was not computed; a segmented scan is required (>16 GB otherwise)." #4374: narrowed the step to reuse #2868's cells; carries R0(29#,30)=0.7233172113, R0(29#,60)=0.4586279069, 23# correction 0.7241584949. #5625: the setter; closed part (c). #2868 (accepted/verified): the reused cells at source. #4387: identity E(h) = Σ_j max(D_j − h, 0), B1/B2 identically trivial once h ≥ G_κ(x#). #4379: the dial's window exceeds the K(h) window by 3.54×–18.86× at every rung and lane.

Why still open. The step's content beyond the record — the 29# T-index at L=30/60, p00(h), exact-product p00_null(h), K(h) with the published R0, the M=199 control on both c lanes, and tests (a) collapse K_x(h) = Phi(h/L) and (b) the c=1 amplitude law over 11#..29# — is computed by no return; both the success clause ("within 5% across 11#..29#") and the failure clause ("K(1) falls inside the permutation band at 29#") are decidable only with the 29# rung. Part (c) is closed by the record and excluded by the step's own text. The step is already the reuse-narrowed form, so verbatim copy orders no recomputation of published cells. Nothing recorded after #5625 answers it or any part of it beyond what it already reuses → outcome promising, next_step = the served step copied exactly.

Not claimed. No p00, K(h) or z value at 29#; no G2/beta_2/twin-prime bound; the K(h)→M_2k(h) transfer stays conjectural and closed at the dial's window; no route-232 experiment run.

  1. Custody (checker C1/C2). GET /research-routes/232 -> state active, revision 3, last_return_id 4374, obstacle null; jobs 11382 assigned / 9003 expired / 8998, 5363 returned / 5365 expired; basis exactly #2570, #2572, #4374. Canonical next_step sha256 (sorted keys, compact separators, UTF-8) aff11570...d29ed85 = #4374.next_step_sha256 = the brief's step object. The step has not moved since #4374.

  2. Absence (C3). 28 returns fetched and text-scanned. Outside the route's own set no return reports a numbered K(.)/p00(.) value and none carries PREREGISTRATION_fq.md; K(h) elsewhere occurs only in step checks that quote or exclude it: #4379, #4387, #4682, #4700, #4767, #5114. Set: brief candidates 5608/5604/5602/5596/5574/5569/5551/5542/5535/5517/5508/5502; #4374's citers 4377/4379/4387/4551/4682/4700/4767/5114; linked-route latest 4383(143)/5130(199)/4783(200)/4762(201)/4713(216).

  3. Part (c), identity (C5). #4387: for a q-periodic word with cyclic site gaps D_j, E(h) = sum_j max(D_j - h, 0); re-brute-forced here on 60 synthetic cyclic words (n in {7,11,13,17,23}), every h: 0 mismatches against a direct window scan.

  4. Part (c), arithmetic (C4). E(h)=0 iff h >= max_j D_j = G2(x#). Dial window h >= x^2 (#4379; route 143's own cells (17,289),(19,361) per #4383). G2 = 108/150/204 at 17#/19#/23# (served G2-STATE, quoted by #5502), 258 at 29# (#5502; lane PROV #3439 for 17#). Ratios x^2/G2 = 2.676, 2.407, 2.593, 3.260 -> E(h)=0 at every rung, so no K(h)-based tail bound can supply route 143's M_2k(h) input there.

  5. Part (c), window mismatch (C6). Recomputation: q = 2prod_{3<=p<=x} p, K = prod_{3<=p<=x}(p-2), L = round(cq/K) -> c=1: 17,20,23,26,28; c=2: 34,40,46,51,56 (11#..23#). x^2/L in [3.559 (11#,c=2), 18.893 (23#,c=1)]; #4379 quotes 3.54-18.86.

  6. Compared returns. #4374: served step, narrows to reuse #2868's 29# cells. #4379: K(h) vs E(h)/M_2k(h) object mismatch, window ratio, step "still open" at 03:46Z. #4387: the identity, B1 tautology, vacuity once h >= G_kappa. #4383: route 143's (17,289),(19,361) cells. #4682 (216): 31# rung open, cites K(h) to exclude. #4762 (201): R0_fix at 29# - scalar deficit, not the pair profile. #4713 (216): R_A/skew/kurtosis at 29#/31#. #4783 (200): count-tail negative, ratio-level stays open. #5130 (199): windowed cumulants. #4700/#4767/#5114/#4551: other routes' step checks. #5502 (363): served G2 ladder. The twelve brief candidates on 224/293/363 are CRT-parent-word, two-interaction and lift-defect work.

  7. Checker: check_b62.py, stdlib offline over this run's snapshots -> 20/20 PASS exit 0; --corrupt (G2(29#)->1000) exit 1 on C4.29/C4.5.

  8. Not claimed: no p00/K(h)/z/Holm/collapse value at 29#; no bound on G2, beta_2 or twin primes; E(h)=0 is about A_q at h >= x^2 only.

Comparison only: no experiment run, no computation reproduced (the step-check rule). Route record rev 2 (active, last_return_id 2572, obstacle null); held step method sha256 008d4433…65f86, canonical object f20ff02a…6f91 ; the brief's step is object-equal.

What each return settles

Why the step is still open

No return computes p00(h), K(h) or z/Holm at 29#, and none tests the six-rung collapse K_x(h)=Φ(h/L) or the c=1 amplitude law on 11#..29#. The success/failure clauses (collapse within 5% or fixed per-prime factor; else K(1) inside the permutation band at 29# closes the route) remain decidable only by the 29# computation.

Why progress, not promising-verbatim

#2868 already publishes the exact scalar cells the frozen rule's application at x=29 produces before any pair scan (p0/Pq0, p0_null/Pnull, R0 at L=30 and 60, plus the 23# c=1 R0 correction). Verbatim copy would direct recomputation of published cells, which the assignment forbids; the replacement step reuses them and asks only for the missing content: the T-index pair scan at 29#, K(h) with the published R0, the M=199 control, and the unchanged six-rung tests (a),(b),(c). This mirrors #2874's identical decision.

Independent check

check_ni.py (stdlib, offline, over served snapshots): 101/101 PASS, exit 0; --corrupt 5/10/21/27 each exit 1. Exact-product nulls at 29# (L=30,60) and 23# (L=28,56) recomputed from (q,K,L) match the served values; R0(23#,28)=0.7241584949 re-derived from #2491's zA via #2874.

Coordination and limits

Message #7247 (job 8988's step check, quoted in the brief; messages endpoint 404s on direct fetch): K(h) sits at the wheel scale (L 17..56), not the dial's h — bears on (c)'s framing, not a substitute for the test. The absence screen covers the fetched records only; #2868's 29# cells are reused read-only. No claim about G2, beta_2 or twin primes.

All numbers below are re-derived from this run's own files by an independent checker that does not import the producer: check_hs.py (stdlib + numpy, offline). Result: checks_ok = 1317, checks_fail = 0, exit 0; --corrupt (planted +0.05 shift in every K(h)) → 161 fail, exit 1.

Rule reuse is validated against an external source

scan_x17.json reproduces the served return #2570 table at x=17 exactly: q=510510, K=22275, R0=0.704495 (c=1) and 0.304046 (c=2), and every K(h) to <1e-9. The served cluster_fq.json was fetched read-only from GET /files/&lt;sha&gt; for #2570 (25 files, parent recorded in served/return_2570.json) — a genuinely external copy, not a sibling run directory.

The checker additionally re-derives the small rungs from scratch by a different method: A_q = {a : gcd(a(a+2),q)=1} by direct math.gcd enumeration, naive O(q·L) window counts, and T = {t : N_t = 0} with p00(h) = #{(t,t+h) ⊂ T}/q. All p0/p00/K(h)/R0 match to <1e-9.

Exact null

p0_null and p00_null(h) use the frozen form C(q-U,K)/C(q,K); the implementation uses the algebraically identical product ratio ∏_{i&lt;U}(q-K-i)/(q-i) (the math.comb form has ~5e5-digit integers at 19# and stalled >600 s). The identity is verified at x=11,13,17 in check_hs.py against Fraction(math.comb(q-U,K), math.comb(q,K)) for every (L, h).

Structural checks

For each rung: q = ∏_{p≤x} p, K = ∏_{2&lt;p≤x}(p-2) (also equals the brute-force |A_q|), mg = q/K, L = round(c·mg), hmax = min(L, floor(q/8)), 2L ≤ q; p0·q is an integer; p00(h) ≤ p0; R0 = p0/p0_null.

Control / z / Holm / sham guard

The control is the frozen one: M = 199 size-matched draws (independent thinning to exactly K positions), the same K(h) formula per draw, z_obs(h) = (K(h)−mean_h)/sd_h, p_two = erfc(|z|/√2), plus one independent sham draw per cell as the calibration guard.

What the evidence does NOT show

No G2, beta_2 or twin-prime-infinity bound; the K(h) → M_2k(h) transfer is conjectural; the wheel diagnostic is a diagnostic; 29# was not computed. See report_hs.md § Honest limits.