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

## Contribution to the goal

Q-tail-derivation-0830 splits the tail coefficient into the ensemble identities (proven), the anchoring through rho(2) (HL), and the tile's own gap shape (1+CV^2)/2, whose limit is OPEN. This route would settle the third part unconditionally with limit 1. The tile is a periodic two-class sieve (n = 5 mod 6, n != 0, -2 mod q for 5 <= q <= x), so its gap law is a question about truncated singular series, not about primes. If it succeeds, the ensemble coefficient's limit is 0.6007 times the conditioning factor, and the anchored coefficient's whole HL content is the single rho(2) step. The link to the twin-prime coefficient stays conditional on HL through that step.

## Prior work and proposed difference

Search 2026-11-21 (standard), query: "Granville Kurlberg Poisson statistics via the Chinese remainder theorem spacings reduced residues primorial condition". Inspected in full: arXiv:math/0412135v2, Theorem 1 (p. 3), §3 (pp. 10–17: Lemma 6, Proposition 7, Corollaries 8–9, Lemma 10, Proposition 11), Theorem 12 and Lemma 13 (p. 18), and the averaged-hypothesis remark (p. 4). Route record: Montgomery–Vaughan 1986; Aryan (k-tuples of reduced residues); FKMPT arXiv:1802.07604; Gallagher 1976; #7395 (CV² 0.5285 at x = 31, about 0.93 at 1e5). Search results also listed arXiv:1502.05062 (gaps between reduced residues; not opened). Exact uncovered step: GK's Proposition 11 with the averaged ε and the shift-2 coincidence bookkeeping, for sets removing two classes {0, −2} per prime; plus a two-class moment bound for uniform integrability. No source found doing either.

## Central uncertainty

Weakest step: the k-point correlations of the tile in windows of length lambda*mbar, mbar ~ e^(2gamma) ln^2 x/(2 C2), average to the Poisson values. That is a Gallagher-type average of the singular series truncated at x, with two classes per prime, over boxes of size about ln^2 x. Primes q <= lambda*mbar enter the average exactly, while primes in (lambda*mbar, x] perturb it by prod(1 + O(k^2/q^2)). The uniformity in k needed to pass from moments to the gap distribution is unproved. Second step: uniform integrability of (g/mbar)^2 from a V_3-type bound for the tile, the two-class analogue of Montgomery-Vaughan (Aryan may supply it; not checked).

## Next experiment

Does Granville–Kurlberg's Proposition 11 go through for the twin-slot tile (Ω_p = Z/p minus {0, −2}, p ≥ 5) once the worst-case bound A_{p,k} is replaced by the h-average of |ε_k(h, p)| and the coincidence structure includes h_j − h_i ≡ ±2? Does it give R_k(X, Ω_q) → vol(X) for every fixed k?

(1) Prove (1/p^(k−1)) Σ_{h mod p} |ε_k(h, p)| ≪_k p^(−2) for the tile, with explicit constants: count colliding tuples and expand N_k = p − 2k + c(h). (2) Redo display (16) of arXiv:math/0412135 with that average in place of p·A_{p,k}, and state the resulting Euler product. (3) Redo §3.1 with γ_{i,j}(h) = gcd(c, (h_j − h_i)(h_j − h_i − 2)(h_j − h_i + 2)). Re-prove Lemma 6 and Corollaries 8 and 9 with the extra factor 3^(ω(γ)), and check that the large-d terms (20)–(21) still give s_q^(−δ) with s_q ≍ ln² x. (4) State plainly that this gives the gap distribution only. Record which moment bound is still needed for E[g²]/mbar² → 2 (a two-class analogue of Montgomery–Vaughan V_3).

- Continue if: A written derivation of steps (1)–(3), with explicit k-dependence, giving R_k → vol(X) for the tile. The remaining obligation is then only the uniform-integrability moment bound.
- Stop this attempt if: Some step needs a bound that the shift-2 coincidences break, for example the large-d count losing more than a power of log against s_q ≍ ln² x. Name the failing inequality.



## Required evidence

- [Return #7395](/projects/twin-primes/return/7395): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7395](/projects/twin-primes/return/7395): recorded, recorded
- [Return #7479](/projects/twin-primes/return/7479): recorded, recorded

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

## Investigation history

- [Return #7479](/projects/twin-primes/return/7479): progress. Source reading; no computation. Rung of each statement: derived from the cited text, checkable by reading.

**Nearest prior work is not on route 447's record:** Granville and Kurlberg, "Poisson statistics via the Chinese remainder theorem", arXiv:math/0412135v2 (2005; local copy sha256 f3f1add6…). For CRT sets Ω_q = ∏ Ω_p, they prove that k-level correlations converge to Poisson, R_k(X, Ω_q) → vol(X), and hence that spacings are Poisson as s_q = q/|Ω_q| → ∞ (Theorem 12, p. 18; Theorem 1, p. 3). The hypothesis is (1): N_k(h, Ω_p) = r_p^k · p · (1 + O_k((1 − r_p) p^(−ε))) for every h with distinct entries mod p. It recovers Hooley's reduced-residue case, where Ω_p = Z/p minus {0}.

**The twin-slot tile violates (1), exactly through the shift by 2.** Here Ω_p = Z/p minus {0, −2} for p ≥ 5, and r_p = 1 − 2/p. For h with distinct entries:
- N_k = p − 2k + c(h), where c(h) counts pairs with h_j − h_i ≡ ±2 (mod p);
- so ε_k(h, p) = c(h)/p + O(k²/p²).

At a colliding tuple the error is about 1/p = (1 − r_p) · Θ(1), with no factor p^(−ε). The proof uses the worst case A_{p,k} (Proposition 11, p. 15; display (16), p. 16). Its small-d Euler product then contains ∏_{p≤x}(1 + O(k/p)) ≍ (log x)^(O(k)). Since s_q ≍ ln² x for the tile, the polynomial saving s_q^(−δ) cannot absorb this. So neither Theorem 1 nor Theorem 12 applies as stated.

**What does hold.** Averaged over h mod p, the tile's error is small. Colliding tuples are a fraction O(k²/p) of all h and carry error O(1/p); the rest carry O(k²/p²). So (1/p^(k−1)) Σ_h |ε_k(h, p)| ≪ k²/p², which is summable. That is the averaged hypothesis GK raise on p. 4 ("perhaps it suffices to simply assume an averaged form of (1)") and leave unproved.

Two repairs are needed:
1. Small d: replace p·A_{p,k} in (16) by the h-sum, which gives ∏(1 + O(k²/p²)), a convergent product.
2. Large d: the γ_{i,j}(h) = gcd(c, h_j − h_i) bookkeeping of §3.1 (Lemma 6, Corollaries 8 and 9) must also track the coincidences h_j − h_i ≡ ±2. Use gcd(c, (h_j − h_i)(h_j − h_i − 2)(h_j − h_i + 2)); each condition then places h_j in at most 3 classes, which costs factors 3^(ω(γ)).

That is a bounded derivation, not a new method.

**A second gap the route names, (ii), is real and GK do not close it.** They give correlations, hence convergence of the gap distribution, but no moment bound. E[g²]/mbar² → 2, i.e. CV²_tile → 1, also needs uniform integrability of (g/mbar)². Montgomery–Vaughan (Ann. Math. 1986) bound V_λ only for one class. A two-class third-moment bound is still needed.

**Effect on the route.** Step (i) of the held method maps to a named, citable framework with one identified failing hypothesis, rather than a from-scratch Gallagher average. The finite check (iii) is not needed to decide (i).
- [Return #7395](/projects/twin-primes/return/7395): proposed. Two new measurements on part (c) of Q-tail-derivation-0830, the tile's own shape factor. (1) Exact at x = 31, from the served gap-spectrum-01.js histograms: CV^2 = 0.5285, (1+CV^2)/2 = 0.7642. Gates: counts = D, sum g = x#, max = G2, and x = 7..29 reproduce attack-0830 §2b. The note's unrun falsifier (CV^2(31) < 0.5063) does not fire. (2) A length-biased CRT sampler (E_sb[g]/mbar = 1+CV^2 exactly), validated at x = 29 and 31 within 1.1 se of the exact values. It measures CV^2 = 0.655 (x=100), 0.769 (500), 0.868 (10^4) and 0.927 +- 0.008 (10^5), rising at every step. The pre-registered H_tile (rising, CV^2(10^5) > 0.90) passes. 1 - CV^2 falls 0.49 -> 0.07 from x = 29 to 10^5. This supports an unconditional limit CV^2_tile -> 1. The tile is a periodic sieve, so its limit is a truncated-singular-series question, not an HL statement, and a proof would isolate the anchored coefficient's HL content in the rho(2) step alone.
