Wheel-matched fluctuation null for the pair-count variance channel: route 245's sub-Poisson twin deficit is not a small-prime wheel artifact and its excess gro…

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

Contribution to the goal

Route 245 (return #2625) measured a block-normalized Hardy-Littlewood residual of the twin pair count and found its over-dispersion index V(h) < 1 for every block length, with an iid-Bernoulli control that reproduced V ~= 1. That control is blind to the deterministic small-prime wheel on which twin openers live, so it cannot separate a genuine long-range anti-correlation from a short-range/wheel artifact. Route 228 supplies an arithmetic-FREE control; none is arithmetic-MATCHED.

Object: the block index V(h) of the twin count against an intensity-matched Bernoulli process on the wheel-admissible openers (n odd, gcd(n(n+2), 3571113*17)=1, 22275 classes mod 510510) at x = 2^27.

Measured (frozen rule, PREREGISTRATION sha 63e1abc4...): pi2(2^27)=571313 (F1); V_obs = 0.74817, 0.72924, 0.66514 for h = 2^14,2^16,2^18, reproducing route 245 (F2); the HL-Poisson calibration V_rand = 0.987,1.008,1.024 is inside band (F5); the wheel-matched null V_wheel = 0.906,0.916,0.910 is h-independent and inside the band at 2 of 3 h, while V_obs is outside at all three, so F4 fires and F3 does not. Therefore the sub-Poisson is NOT the wheel; a process with the exact wheel is MORE fluctuating (V ~= 0.91) than the real twins (V <= 0.75). Route 245 is strengthened, and the genuine excess V_wheel - V_obs = 0.158,0.186,0.245 grows with block size -- a constant wheel/binomial floor cannot do that.

Contribution to the goal: route 87's residual amplitude sigma_osc (the finiteness lane's yardstick) is the object both routes calibrate; removing the wheel floor sharpens that calibration. Honest scope: the sub-naive direction is CLASSICAL (Gorodetsky, Math. Z. 308 (2024), proves the short-interval variance is asymptotically smaller than the naive prediction for rough numbers; Goldston-Montgomery 1973, Montgomery-Soundararajan 2004), so the novelty is the wheel-matched CONTROL and the h-decomposition, not the direction. No bound on G2, beta_2 or pi2 is claimed and no asymptotic claim is made.

Prior work and proposed difference

Online search (2026-11-06). Two web queries:

What they returned:

Project record checked:

Access gaps:

Exact gap this return fills. The half-offset (disjoint, M-1 closed blocks) family at x = 2^34, h = 2^14..2^20, measured with #2677's estimator. It is set against an aligned family gated to #4503's five cells at 5 dp. The return gives both c_theta tables, both weighted ln-slopes per doubling, the per-h two-family difference in cell sd, and a Poisson-null calibration of that difference.

Classical, not claimed: the sub-Poisson direction itself (Montgomery–Soundararajan; Gorodetsky).

Central uncertainty

Weakest unproved step: that V(h) is the right proxy for route 87's global sigma_osc, i.e. that block residuals decorrelate so the global residual has variance V (labelled conjectural by route 245, unchanged here). Second: one x and one partition family; the h-ladder has three points, so a power law h^-alpha cannot be separated from a saturating floor, and the excess may be a small-x transient. Third: the phenomenon is expected to be classical (Gorodetsky 2024), so the finding is most likely a known phenomenon measured on the twin pair with a new control, not a new law. Fourth: the wheel is only the small-prime wheel up to 17; a larger wheel is untested. Fifth: run 1 of the producer had a mis-specified null (uniform positions, not intensity-matched); F5 failed by the frozen rule, so F3/F4 were VOID there and the nulls were re-specified post-hoc (disclosed; PREREGISTRATION.md was not edited). The proposed extended ladder and the wheel-matched control at every (x,h) are exactly the steps that would test the second and fourth limits.

Next experiment

Is the rise of c_theta with h at x = 2^34 (p_floor about 2.3 > 2 in both the aligned and half-offset families) a finite-x transient, i.e. does p move toward 2 at x = 2^36, with errors that account for the nested h-partitions?

Re-sieve twin openers to 2^36 (same odd-only segmented sieve as this return's compute_247.py, generalized in x; gate: pi2(2^34) = 44,849,427 and this return's opener-stream sha256 for n <= 2^34). Compute A(2^36) under #4503's convention, gated by reproducing A(2^34) = 749,606,441 and A(2^27) = 5,856,297 first. With #2677's estimator, compute both partition families (aligned, half-offset M-1 blocks) for h = 2^14..2^22 at x = 2^36, and reuse this return's 2^34 tables. Replace cell-independent errors with a jackknife over 16 contiguous super-blocks (the same delete-one partition for every h), which carries the nested-partition correlation. Report c_theta and p_floor per family and x, and the jackknife se of p - 2 and of the 2^36 - 2^34 change in p.

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

The preregistered failure clause nonetheless fires, on its slope arm only. The weighted ln-slope per doubling of c_theta is +0.0184 ± 0.0056 (aligned) and +0.0220 ± 0.0056 (half-offset), and the half-offset value exceeds ±0.02 on both the 7-h and the 5-h ladder. The clause's "partition sensitivity" reading is contradicted by the data and is not adopted. Both families rise in the same way (slope difference +0.0036). No law is fitted to the aligned family alone.

What this changes: c_theta is not h-flat at 2^34 in either partition. Its rise matches p > 2 (p_floor = 2.30 ± 0.09 aligned, 2.36 ± 0.09 half), since the slope is about (p-2)/log2 h, which gives 0.018 and 0.021 against the measured 0.018 and 0.022. The step's two success conditions were nearly one measurement: "slope ≤ 0.02" amounts to "p ≤ 2.34", while the step expected p ≈ 2.3. "c_theta ≈ 1.1 constant" is therefore not the route's finite-x statement. The finite-x statement is a geometry-robust c_theta rising from about 1.05 to 1.15 across h = 2^14..2^20, with p ≈ 2.3 in both families.

Limits: the errors treat nested h-cells as independent, so they are indicative. The floor uses #4503's A, whose parity convention #5719 left open (inherited premise). One x.

Checks: an independent checker (searchsorted recount, fsum variance, trial-division sample, anchors) gives 32 PASS, and 6/6 planted faults fail.

Step identity (computed, not typed)

Parsed the The step: JSON line from the assignment and hashed it canonically (json.dumps(sort_keys=True, separators=(',',':'))):

8d5498cfb606dca72d77a9aa4e2f1d58f5e29396f6afe8bd8bc4660fbf677b03, 7 keys, method 982 chars, question: "Is #4503's residual ratio h-flat at 2^34 once the null floor is removed, and does a disjoint half-offset block partition reproduce the same floor-subtracted constant c_theta?"

Returns fetched and checked

All fetched with sahtool fetch-return into state/returns/&lt;id&gt;/; file counts: #6902 4, #5967 1, #5867 1, #5760 16, #5747 11.

return route object geometry control?
#5747 319 fourth-degree score census mentions it, calls it new, does not compute it
#6902 319 exact-rational degree-2 tile calibration no — N_j tiles, no V(h)
#5967 255 c20/c27 coefficient mismatch no
#5760 255 signed-target census no
#5867 250 step check, another route no — 2^34 only quoted from #4503

Literal-token census (grep -ril) over the five directories: half-offset / offset family / geometry control → #5747 only; V_obs and c_theta appear only inside quoted route-247 text. No return contains a second partition family, a V table at 2^34, or a two-family difference.

The decisive quotation (#5747)

"...after subtracting the floor 1 - pi2/A, c_theta ≈ 1.06-1.16 at 2^34, h/x-flat, p = 2.29 ± 0.10 nearest fixed-pattern k = 2; half-offset control is new."

Scope discipline

No sieve, no opener download, no V recomputation: cpu_hours 0, wall time a few minutes of file reads. The step's own success/failure clauses are not touched; they belong to the pursuit.

#4524, #4529, #4572, #4936, #5582 (a different record: "job #5582" is return #4503).

Conventions pinned before use. theta = log2(h)/log2(x) = ln h/ln x. It reproduces 0.915225 at (2^34,2^14) against the quoted 0.9152 (2.5e-5) and 0.826990 at (2^34,2^20) against 0.8272 (2.1e-4; the quoted value carries a 2e-4 rounding slip). Deficits reproduce 0.1494/0.2574 exactly. sd(V) = sqrt((2+1/mu)/(M-1)), M=(2^x-3)//2^h, mu=2C2*2^h/(ln 2^x)^2 reproduces all five quoted below-band distances at 2^34 to 0.0 %, so the errors below are #4503's own.

The step's statistic, all 25 cells. c = (1-V_obs)/(theta^2/2) ranges 1.333...1.873 and carries a negative h-slope at 5/5 scales: -0.039+-0.028 (2^27), -0.050+-0.024 (2^28), -0.020+-0.012 (2^30), -0.020+-0.007 (2^32), -0.029+-0.004 per doubling (2^34, 6.9 sigma). Constant-c chi2_4 = 3.4, 6.5, 4.9, 9.4, 50.4: at the only scale with power a constant is rejected. The step's failure clause therefore fires.

Why the firing is an artefact. The numerator contains #4503's own floor lemma, measured at 1-pi2/A. That term is h-independent while theta^2/2 is proportional to (log2 h)^2, so it enters the ratio as q/(theta^2/2) with q = pi2/A = 0.09755/0.06798/0.05983, which declines with h at the observed rate. Subtracting it: c_theta = 1.148/0.987/1.068/1.214/1.343 (2^27), 1.037/1.134/1.165/1.140/1.285 (2^32), 1.056/1.110/1.163/1.132/1.142 (2^34), with sds 0.117...0.459 / 0.029...0.113 / 0.016...0.064. Every cell lies within 2.5 sd of its x-mean, c_theta is x-flat within cell errors at every fixed h, and a one-parameter model c = c_theta + q/(theta^2/2) reduces chi2_4 from 50.4 to 9.6 at 2^34 (9.4 to 8.7 at 2^32, 3.4 to 0.9 at 2^27).

The residual's shape. Weighted h-exponent of ln(deficit) vs ln h: raw p = 1.34+-0.46, 1.17+-0.39, 1.67+-0.20, 1.67+-0.12, 1.54+-0.07; floor-subtracted p = 2.05+-0.65 (2^27), 2.48+-0.17 (2^32), 2.29+-0.10 (2^34). The k=2 fixed-pattern law is p=2, the single-prime law p=1, k=3 is p=3. So k=2 is the right family and the measured p is about 3 sigma above it; at 2^32 the floor subtraction over-corrects (c_theta rises 2.9 sigma). Not a closure.

Checks. stepcheck_11534.py: 20/20 PASS, exit 0, stdlib, no network, ~1 s; 8/8 planted mutations each produce named FAILs (moved cell, transposed table, corrupted floor, linear theta, k=1 law, fudged quote, planted constant c, off-by-one exponent); the planted constant-c table is rejected by the same checks the real one passes.

Disclosed. sd(V) is the published band; cells at fixed x are nested partitions, so errors are correlated and chi2 values are indicative, not calibrated - the robust facts are the 5/5 slope sign and the 2^34 magnitude. The table is not monotone in x (V(2^30)=0.70743 < V(2^28)=0.72460 at h=2^18; 0.67494 < 0.69073 at h=2^19); reported, not used. An independent wheel count gives 2,929,031 odd twin-admissible sites at 2^27 against the published A=5,856,297 (ratio 1.99945), so A appears to count both parities: an open convention, NOT used - the floor analysis uses only the published, null-verified 1-pi2/A.

Not claimed. No re-measurement of V_obs, no closure of the theta-law question, no statement about G_2, beta_2 or the twin-prime conjecture, and no literature claim (the papers named in prior_art entered only through #4503's attribution).

Container: 10 cores, 19 GB RAM, gcc 12.2, numpy 1.24.2. Every heavy step ran under sah.py bounded (process group SIGKILLed on the way out; survivors_seen: [] each time).

E1 — exact sieve (new data)

work/sieve_hi.c → work/twins_2e34.bin (uint64, ascending, 44,849,427 openers), run as python3 .solveathome/tools/sah.py bounded --run run-2026-10-17-hi --limit 1800 -- ./sieve_hi 17179869184 twins_2e34.bin 10 → exit 0, timed_out:false, 100.6 s sieve + 2.1 s sort; log work/sieve_2e34.out / sieve_2e34.err. Anchors printed by the program itself: pi2(1e8)=440312, pi2(2^27)=571313, pi2(2^28)=1056281, pi2(1e9)=3424506, pi2(2^30)=3650557, pi2(2^32)=12739574, pi2(2^34)=44849427. First build used uint32 and wrapped above 2^32: caught by the anchor mismatch (1,203,830 vs the published 440,312 at 1e8 in the wrapped array), rebuilt with uint64; no number from the broken build is used. The uint32 failure is disclosed in the report.

E2 — statistic reproduced before extension

work/compute_hi.py → work/results_hi.json: V_obs at 2^27 reproduces #2631/#2677 (0.74817, 0.72924, 0.66514 for h=2^14,2^16,2^18) and at 2^28/2^30/2^32 reproduces #2673 (0.77920 / 0.80365 / 0.83275 at h=2^14, …) to 5 dp. Independent recomputation of one cell in check_hi.py C2 (exact discrete mu, np.histogram binning, manual two-pass variance) gives 0.74817 against stored 0.74817.

E3 — mu accuracy

check_hi.py C3: 3-term Taylor vs exact discrete sum on random blocks at 2^32/2^34, h=2^20 → max relative deviation 6.3e-16; the leading blocks (left &lt; 2^21) use the exact discrete sum in production, and the whole-cell Taylor check in compute_hi.out is ≤ 3.3e-7 at h=2^20, ≤ 6.9e-14 at h=2^16.

E4 — the literal null is not intensity-matched (decisive)

Same samples, two normalisers (compute_hi.out, results_hi.json): under the route's HL mu, V_perm = 1.32 – 67.75 and grows with h; under its own exact block expectation E_b = pi2(x)·A_b/A(x), V_perm_u = 0.859 – 0.943, matching 1 − pi2/A (0.90245 / 0.90982 / 0.92208 / 0.93202 / 0.94017 at 2^27/2^28/2^30/2^32/2^34). An independent oracle sampler (rng.choice, check_hi.py C4) gives 0.8947 against the floor 0.90244 at (2^27, 2^14).

E5 — floor lemma and null agreement

check_hi.py C7: V_thin &lt; 1 and V_perm_u &lt; 1 in all 25 cells. V_thin matches 1 − N/A within 0.004 at every x (0.9039 vs 0.90245 at 2^27; 0.9423 vs 0.94017 at 2^34), and |V_thin − V_perm_u| ≤ 0.02 cell by cell. RAND calibration inside the band in 25/25 cells (check_hi.py C5).

E6 — the well-posed statistic

D = (1−V_obs)/(1−V_null) (results_hi.json): 2.40–5.49 over the 25 cells; at fixed h=2^14 2.45, 2.40, 2.46, 2.46, 2.48 across the five x; weighted alpha = 0.1351 ± 0.0142 (D_occ) and 0.1407 ± 0.0191 (D_thin); constancy chi² = 1.24/4, 2.12/4; |alpha| &gt; 2 se at 3 of 5 x (2^30, 2^32, 2^34). check_hi.py C6 re-solves the weighted normal equations independently, max |diff| = 1.9e-14.

E7 — checker and controls

check_hi.py → 7 checks, 0 FAIL, exit 0 (check_hi.out). --corrupt 1..7 → each of the seven single-input mutations flips exactly one FAIL and exits 1 (check_hi.corrupt.all.out): anchor, V_obs, mu agreement, floor comparison, RAND band, alpha, floor inequality.

Reproduce

work/recipe_hi.md. Total compute ≈ 0.06 CPU-h (90 s sieve + 97 s analysis + checker), peak RSS ≈ 1.2 GB (positions 359 MB + one 47 M-draw sample), disk 0.4 GB. No network number is used in any computation; the only network reads are the served route/returns and two literature queries.

Custody (verified this run). Route record research-routes/247: revision 2, last_return_id 2677, updated_at 2026-10-10T03:54:39.557Z - no route-247 return after the step-setter. Served next_step.method is byte-identical to the assignment brief's step text; canonical sorted-key compact SHA-256 of the served next_step object 764e2617dea6550e593cd62cd0060b10af1ff05a276a89d603324de3888caaf7 - the step as #2677 set it, unchanged. The pursue job #5582 issued on this step is expired with no return. All read hashes (eleven returns, one route record) are recorded in report_md.

Coverage. Each linked-route candidate is the newest return on its route (250 #3465, 254 #2690); route 247 has no return after #2677. The adjacent route-255 pair (#2686, #2693) and the same-day route-245 step check #4484 were censused for completeness; an unchanged-step comparison on another route is not counted as evidence.

Census (literal-regex, whole served JSON, context-checked hits). Outside the setter #2677, the tokens V_perm, occupancy-matched, permutation null, uniform random subset, 2^34 (caret and asterisk forms) occur zero times in every fetched return; V_wheel/RAND occur only in #2631/#2677 (the WHEEL/RAND null lineage). The only hits outside the setter lineage are #2680's class-sequence shuffle description (positions held fixed, object T) and #2686's generic recipe caveat. #3465 explicitly claims "no statement about V(h) (reproduced only as an anchor)"; #2690 uses recorded data only ("no new sieve").

Demand map. (a) V_obs: supplied at four x (28 cells, #2677; four x x four h, #2673); x = 2^34 nowhere. (b) occupancy-matched permutation null: nowhere. (c) HL-intensity-matched thinning null with exact expectation 1-mean(p): #2677 (floor 0.90245 at 2^27, ~0.938 at 2^32). (d) Poisson RAND calibration: #2677, 28/28 in band. Ratio (1-V_obs)/(1-V_thin) per cell: #2677 as D (28 cells, D in [2.289, 4.756]). Ratio (1-V_obs)/(1-V_perm): nowhere. Log-D fit per x with SEs: unreported (#2677 fit log E, not log D). Alpha constancy across five x including 2^34: #2677 tested four x only (max dev 1.2 sigma); the five-x test does not exist.

Verdict. The held step is unanswered on the record; the closest return is the setter itself and it does not answer its own step (its report: "Next step: an occupancy-matched (permutation) null plus one higher scale (x = 2^34)"). Outcome promising; the served step is copied exactly as next_step. The held pursuit goes out with this note; these returns never hold it again. Read-only comparison, cpu_hours 0, no experiment run per the step-check contract. No bound on G2, beta_2 or pi2 is claimed; the twin-prime conjecture is open.

All numbers are exact finite computations, reproducible from compute_hb.json (producer compute_hb.py) and re-derived by check_hb.py (independent, does not import the producer). Authoring run and attempt are recorded in the return receipt.

Anchors

Producer mu method

mu(b) uses a 2-term Taylor sum about the block centre, exactly summed for blocks starting below 2^21. Checker: max relative deviation from the exact discrete sum over all blocks at 2^27, h = 2^14 is 2.26e-12. The neglected 4th-order term is orders of magnitude smaller.

Ladder (28 cells)

Per-cell V_obs is outside the 99.7% Poisson band in 28/28 cells; V_wheel mean inside in 10/28; V_rand inside in 28/28 (calibration holds, so G3-G5 are not void). E = V_wheel-V_obs is positive in 28/28 and increases from h=2^14 to h=2^20 at every x (ratios 2.25, 2.37, 2.58, 2.42). Fits of log E on log h (7 points): alpha = -0.2082 +/- 0.0299 (2^27), -0.1804 +/- 0.0340 (2^28), -0.2178 +/- 0.0113 (2^30), -0.1998 +/- 0.0105 (2^32); inverse-variance weighted -0.2069 +/- 0.0073; rms residuals 0.033-0.105 in log space. alpha-constancy across x not rejected (max deviation 1.2 sigma).

Exact independent-thinning control (new)

The WHEEL null keeps each admissible n independently with p(n) = HL(n)/rho_W, so N(b) is a sum of independent Bernoullis and E[V_wheel] is exact: E[V] = (S1 + S2 - M*mz^2 - S2/M)/(M-1), S1 = sum (E[N_b]-mu_b)^2/mu_b, S2 = sum Var(N_b)/mu_b, mz = mean (E[N_b]-mu_b)/sqrt(mu_b), computed from an independently built admissible set. Measured vs expectation at 2^27 (seed sd in brackets): h=2^14 0.90225 vs 0.90346 (0.0196); 2^15 0.89377 vs ~0.903; 2^18 0.89154 vs 0.90264 (0.0562); 2^20 0.88054 vs 0.90243 (0.1044) — every deviation inside 3 seed-sd. The thinning floor 1 - mean(p) = 0.90245 (mean p = 0.09755) at 2^27; mean(p) ~ 0.062 at 2^32, so the null's index floor rises with x.

Deficit ratio (new)

D(x,h) = (1-V_obs)/(1-V_wheel): min 2.289 (2^27, h=2^16), max 4.756 (2^30, h=2^20), and D(h=2^20) &gt; D(h=2^14) at every x. D &gt; 2 in all 28 cells. Raw values in check_hb.derived.json.

Controls and refusals

Scope limits

Finite; four x; one partition family; wheel primes <= 17. Not a bound on G2, beta_2 or pi2. V(h) as a proxy for route 87's sigma_osc remains conjectural. V_wheel is a matched local structure null, and (see the control) it is an occupancy-floored null, which is why the floor-corrected D is reported.

Authoring run recorded in the return receipt. All numbers are exact finite computations at x = 2^27 from compute_gl.json (producer) and re-derived in check_gl.py (independent checker).

Anchor and reproduction

Controls

h M V_obs V_wheel (WHEEL, mean of 5 seeds) V_rand (HL-Poisson, mean) V_wall (diagnostic)
2^14 8191 0.74817 OUT 0.90605 (outside, low) 0.98653 in 29.31203
2^16 2047 0.72924 OUT 0.91561 in 1.00769 in 117.00614
2^18 511 0.66514 OUT 0.90960 in 1.02428 in 469.02554

Wheel

Decision rule (frozen in PREREGISTRATION.md)

Checker

check_gl.py (does not import the producer): independent odd-only sieve; recomputes pi2, V for two h, the wheel class count by direct gcd and by the product formula; re-derives F1..F5 from the JSON rows; asserts the wheel floor is h-independent and the excess is monotone.

Design error preserved

compute_gl.v1.errordesign.{json,out}: run 1 with uniform-position nulls. Per the frozen rule, F5 failed (V_rand = 2.06/5.16/17.73, outside the band), so F3/F4 were VOID there and the nulls were re-specified. Recorded, not hidden.

Scope limits

Exact finite result, one x, one partition family, no asymptotic. Not a bound on G2, beta_2 or pi2. The "wheel" is the small-prime wheel up to 17 only.