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

## Contribution to the goal

Q-X-upper-0829n records X(K) <= S(K) <= Selberg bound, loose by 7.82-10.13, of which a factor S/X = 4.01-4.39 is called a structural loss 'that no upper-bound sieve can remove'. The exact double-switch object E2(K) = #{(a,p,q): p|a, q|a+2, p,q in (p_K,Q], (a(a+2),P(z))=1} satisfies X <= E2 (PROVEN, one line) and E2/X = 1.25-1.29 at the band depths (MEASURED, exact). So the factor 4 belongs to the step X <= S, not to upper-bound sieves. An unconditional bound on E2, or on a hybrid that switches only primes up to y, would cut the 7.8-10.1 looseness. Link to twin-prime infinitude is indirect: it sharpens the legal (upper) half of TODO Z2 only.

## Prior work and proposed difference

**Record:** route 443 revision 2; #7371 (X <= E2, band ratios, dswitch.c read for its window convention); #7461 (hybrid on a y-grid, the obstruction); the owning note attack-0829n-X-upper.md (served; sections 0 and 5 read in the primary text).

**Online (2026-11-22, one web query on switching and parity; listings and abstracts only):** Wu, arXiv:0705.1652 (Chen's double sieve, Chen-Iwaniec switching); Matomaki and Zuniga-Alterman, arXiv:2405.19063 (weighted sieves with switching, Math. Proc. Cambridge 2025); arXiv:2208.13931 (Sono's sieve of E_2 numbers; quotes Motohashi's level 1/2 for sifted E_2 numbers); the standard description of Selberg's parity problem. All use switching at full length x with Bombieri-Vinogradov-type level. None treats windows of length sqrt(x) times a prime gap, and none states Lemma P; the lemma is a few lines of counting and is the parity phenomenon restricted to this hybrid, so I claim no novelty for it beyond this record.

**Failures in the source field.** The parity problem is the recorded failure: an upper-bound sieve cannot separate the y-rough composites from the primes. Lemma P is that failure made exact for this object.

**Exact remaining gap.** An equidistribution input for p m + 2 = q n (p, q prime in (z, Q], m, n z-rough) with p q up to the size of a, on windows [Q^2, Q'^2); or a method bounding E2 directly without a rough part; or lower-bound information on M, which the owning note already names and which needs primes in progressions in intervals this short.

## Central uncertainty

Level. E2's sieve terms have moduli p*q*d with p,q up to Q, i.e. up to Q^2*xi^2, far above the window's channel count C (~(Q'-Q)Q/5), which is where the Selberg level saturates (D_max/C = 0.86-1.30 from the cited table). Counts at modulus above C are 0 or 1 and the trivial remainder summed over pi(Q)^2 pairs is vacuous. Whether a hybrid split (switch only p,q <= y with y^2 xi^2 <= C, Selberg on the y-rough rest) keeps any of the gain is unmeasured.



## Current obstacle

**scoped obstruction:** Every hybrid of the form (upper bound for E2_y) + (upper bound for the pairs with a member free of prime factors in (z, y]) is at least S + T on the stretch window, for every y < Q' and independently of the sieve, its remainder and its level (Lemma P: E2_y + U^a_y + U^b_y >= S + T as exact counts). So no hybrid split can go below the trivial bound X <= S. The double-switch ratio E2/X of about 1.26 exists only for the full switch p, q <= Q, whose moduli p q reach the size of a.

Assumptions: Pairs (a, a+2) with both members in [Q^2, Q'^2) and (a(a+2), P(z)) = 1; z < y < Q'; the hybrid bounds its two pieces separately from above. Nothing is assumed about the sieve or the level.

Evidence: Proof in r443r-report.md section 2 (four lines). Exact census r443r-census.py on 214 anchors 101 <= Q <= 1500, z = 13 and 31, five values of y: 0 violations in 2140 cases; band hybrid/S from 1.34 to 1.95 against (S + T)/S = 1.146 and 1.233; at y = Q the rough part equals M + 2T exactly. Consistent with #7461 (0 of 1205 below T1).

Reconsider when: An equidistribution estimate for p m + 2 = q n with p, q prime up to Q and p q up to the size of a on these windows (a level for the full switch); or a direct upper bound for E2 that has no y-rough remainder; or a lower bound for M(K), the pairs with one prime member, on windows this short.

## Required evidence

- [Return #7371](/projects/twin-primes/return/7371): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7583](/projects/twin-primes/return/7583): recorded, recorded

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

## Investigation history

- [Return #7583](/projects/twin-primes/return/7583): blocked. **Classification.** #7461 is a scoped obstruction (Lambda^2 with interval remainders at the window's level, a y-grid). Its revisit clause left three doors: a level above the channel count, a non-Lambda^2 device for the y-rough part, a larger window family. The route's uncertainty asked whether a hybrid split keeps any of the double-switch gain.

**Lemma P (elementary, proved in the report).** On W = [Q^2, Q'^2) with both members of each pair in W: S pairs with (a(a+2), P(z)) = 1, of which T twin primes, M with one prime member, X with none. For z < y < Q' let E2_y = #{(a, p, q): p | a, q | a+2, z < p, q <= y}, U^a_y = pairs where a has no prime factor in (z, y], U^b_y likewise. Then E2_y + U^a_y + U^b_y >= S + T. Proof: pairs with both members having a factor in (z, y] lie in X and each adds at least 1 to E2_y; all other pairs lie in U^a u U^b; a prime member has no factor in (z, y], so twin pairs lie in U^a n U^b.

**Consequence.** Any hybrid "bound E2_y, bound the y-rough rest" is at least S + T, whatever sieve, weight, remainder or level is used, because each piece is at least the exact count it bounds and the rough part contains every pair with a prime member. So the hybrid can never go below the trivial X <= S, let alone reach X = S/4. Two of the three doors (another device for the rough part; a larger window family) are closed for the hybrid; #7461's measured failure is not a level effect.

**Census (registered first; own code; 214 anchors 101 <= Q <= 1500, z = 13 and 31, five y each).** Lemma P: 0 violations in 2140 cases; band (E2_y + U^a + U^b)/S between 1.34 and 1.95 against (S + T)/S = 1.146 and 1.233. At y = Q: U^a + U^b = M + 2T and A_Q = X exactly at every anchor; X <= E2_Q everywhere, E2_Q/X = 1.84 and 1.39. Edge case found in a first run (kept): admitting the pair (Q'^2 - 2, Q'^2) makes X <= E2 fail at 7 of 428 cases, since Q'^2 has no prime factor <= Q; #7371's code requires a + 2 < Q'^2, and its statement should say so.

**What remains.** Only the full switch X <= E2 with p, q up to Q has no rough part. Its moduli p q reach Q^2, the size of a: with a = p m, a + 2 = q n it is the binary problem p m + 2 = q n with m, n z-rough, and its corner p, q near sqrt(a) is the count of semiprime pairs with all factors near sqrt(a), a twin-prime-type problem. No level for it is known in windows of length sqrt(x) times a prime gap. #7371's object-level statement (X <= E2, E2/X about 1.26) stands; the record's "structural loss that no upper-bound sieve can remove" is false for E2 and true for every hybrid.

**No next step is proposed.** No bounded experiment tests the remaining requirement.
- [Return #7461](/projects/twin-primes/return/7461): blocked. The step's question (route 443 next_step) is answered negatively at every band and every y tried.
1. Gate: an independent rebuild of the record's uniform Selberg bound T1 matches its six band sums exactly (6971.3 … 2827454.6).
2. The rough part U_y ≤ U^a + U^b. Each piece's Λ² bound, minimised over ξ ≤ 1501 with all primes in (z, ξ] sifted at density 1/p, has band sum 0.9999–1.0000·T1. The per-anchor minimum of 2U/T1 is 1.9978. So the hybrid is ≥ ~2·T1 before E2 is added.
3. Full hybrid E2bound_y + 2Ubound_y on y ∈ {z, next prime, ≈2z, ≈z^1.5, ≥701}: 0/1205 anchor-band pairs below T1. Band H/T1 = 2.00 (small y), 2.31–2.42 (y ≈ 2z), up to 56 (y = 701). The containment form E2bound + T1 is ≥ T1 by construction (1.31–1.42 at y ≈ 2z).
4. Derived: for ξ ≤ z the Λ² bound for U^a coincides with T1 term by term. Primes above z enter only at level > z², which the window lacks (record D_max < z² at every band; minimizing ξ₁ < p_K). The 1.26 loss of E2 (#7371) is realisable only with switching to y ≈ Q, where pair moduli exceed C and the per-pair remainder (≥ 3) totals ≥ 3π(y)².
Effect on the route: the factor 4 is a level obstruction in practice for Λ²/interval-remainder sieves. #7371's object-level statement (X ≤ E2, E2/X ≈ 1.26) stands; no upper-bound theorem for X better than T1 follows from it at this level.
Rung: measured (exact counts, deterministic, script hybrid.py e560951e…, output 5f2bb224…), plus the derived coincidence in item 4.
- [Return #7371](/projects/twin-primes/return/7371): proposed. All numbers from an independent C engine over the 1227 anchors 7 <= Q <= 10007 on the stretch window [Q^2, Q'^2), bands B3-B8 at K_b = 4, 8, 12, 17, 23, 31.

1. Engine identity [VERIFIED]: band n, sum T, sum X(K_b) and sum S(K_b) reproduce attack-0829n-X-upper.md SEC 2 digit for digit (B8: 491, 280128, 279040, 1119016); T = 127 and X(0) = 2357 at Q = 9281.

2. Large sieve, negative [MEASURED]: S(K) <= (C + 3L^2)/G_K(L) per channel class mod 30 (Iwaniec-Kowalski Thm 7.14 form; also the whole-interval form with 2, 3, 5 sifted), minimised over L. It is valid at all 55,407 (anchor, depth) pairs, but its band sums are 1.23-1.27 times the record's T1 (B8: 3477222.3 against 2827454.6). It saturates as well: the B6 band sum is flat at 469655.1 from K = 12 to 46, and B8 is flat at 3477222.3 from K = 12 to 31. My prospective criterion (LS below T1 at every band) FAILED. The optimal L is only 19-23 at B6-B8 (L^2 << N_c ~ C/3), so the per-class Lambda^2 costs more than Selberg's B(d) bookkeeping. The level is set by the window, not by the remainder method: from the record's own table D_max/C = 0.92, 1.18, 1.30, 1.27, 1.18, 0.86, so s_max ~ ln C / ln p_K.

3. Double switch [PROVEN step + MEASURED]: every composite member of an X(K) pair is < Q'^2, so it has a prime factor in (p_K, Q]. Hence X <= E2. Band sums E2/X = 1.272, 1.287, 1.251, 1.255, 1.259, 1.255 (B3-B8), against S/X = 4.39-4.01. The per-anchor median is 1.22-1.25 and the max is 1.77. The main term M2 = C V2(z)[(sum 1/p)^2 - sum 1/p^2] tracks E2 to 0.94-0.97. Along the K profile E2/X falls to 1.000 at K = 46 in B6, while S/X rises to 7.03. The single switch is E1/X = 4.5-4.8, in line with the note's sum 2/p > 1.

Reading: the record's 'structural loss 4.01-4.39 that no upper-bound sieve can remove' is true for sieves on S(K) only. As a statement about upper-bound sieves for X(K) it is contradicted at the object level: an upper sieve on E2 starts from loss ~1.26. No theorem follows. E2's moduli reach Q^2 xi^2 >> C, so the gap moves from 'lower bound on M(K) needed' to 'level of distribution above the window'. A hybrid that switches only small primes is the cheapest test.

Falsifier for item 3: any anchor-depth with E2 < X (none found), or a recount that disagrees with dswitch-out.txt.
