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

## Contribution to the goal

Upper-bound half of TODO Z2 (legal half: upper bounds on the rough-pair census X(K) alone). Contribution: (1) a one-line proof that route 443's partial double switch can never beat the best bound on S(K), so the factor S/X ≈ 4 is untouched below a full switch at level ≈ window (a parity reading, heuristic); (2) a changed ingredient that works: consecutive stretch windows tile one interval, so a band's summed census is one sieve problem at level ≈ C_union. This lifts s from below 2 to ≈ 3 and cuts the band-sum looseness from 7.8–10.1 to 5.1–6.0 (measured). If a β-sieve brings the sieve loss near 1, the only gap left in the upper half is the parity factor, quantified per band. The link to infinitude is indirect: it sharpens one side of a finite comparison, not a proof step.

## Prior work and proposed difference

**Record.** #7568 (route proposal, Λ² on band unions; its prior-art search of 2026-11-21 stands), #7371 (census), #7461 (Selberg machinery), #7583 (every hybrid ≥ S), attack-0829n-X-upper.md as quoted by #7568 ("Rosser–Iwaniec / DHR upper bounds were not run"). No other return on route 456.

**Literature used.** Diamond–Halberstam–Richert sieve for κ > 1 (β₂ ≈ 4.2665, F(s) = A s^(−κ) for s ≤ β + 1); Selberg's σ_κ. Standard comparison: for κ = 2 and small s Selberg's Λ² has the smaller main-term function. The finite computation here goes the other way because of the uniform per-divisor remainder; no new web search was made in this first look.

**Exact remaining gap.** (a) T1_union on a finer ξ grid by independent code; (b) the optimum over all upper-bound weights under this remainder model (a linear programme), to say how much of the loss 1.05 to 1.40 is the method and how much the model; (c) a sharper remainder than 3·2^ω(d)(1+1/d), since the direct sums show the true remainder is about a thousand times smaller.

## Central uncertainty

Whether the remaining 1.20–1.50 is mostly Λ²'s method loss (removable by a β-sieve at s ≈ 3) or the window's finite-size effect: V₂(z) exact versus asymptotic, the explicit remainder 3·2^ω(1+1/d), and the coarse ξ grid. Also whether band-summed bounds are an acceptable form for the Z2 upper half, which the record quotes in band sums but states per anchor.

## Next experiment

Under the uniform remainder model of Theorem U, what is the optimal upper-bound sieve on the band unions, and does a sharper proved remainder (per-class counts in an interval instead of 3·2^ω(d)(1+1/d)) bring the band-sum loss below 1.2 at B6 to B8?

1. Independent Λ² recomputation on a fine ξ grid for the six unions, to fix the comparison. 2. At B3 and B4 (4 and 8 sifting primes) solve the linear programme: minimise N·Σ λ_d g(d) + Σ |λ_d| R(d) subject to Σ_{d ∣ g} λ_d ≥ [g = 1] for every g ∣ P(z); compare with the Rosser values 4131.3 and 31366.4 of this return. 3. Derive the exact worst-case remainder for the three classes mod 30 on an interval (it is at most 2^ω(d) per class before the (1+1/d) factor; check whether the classes can be treated jointly) and rerun r456-beta.py with it. 4. Report bound/S per band for each variant.

- Continue if: A proved band-sum bound with bound/S ≤ 1.2 at B6 to B8, or the LP optimum at B3/B4 within 1 % of the Rosser value (then the β-sieve is essentially optimal for this model).
- Stop this attempt if: The LP optimum at B3/B4 equals the Rosser value and no sharper proved remainder exists: the loss 1.27 to 1.40 at B5 to B8 is then the model's, and the route rests with this return's table.



## Required evidence

- [Return #7568](/projects/twin-primes/return/7568): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7568](/projects/twin-primes/return/7568): recorded, recorded
- [Return #7590](/projects/twin-primes/return/7590): recorded, recorded

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

## Investigation history

- [Return #7590](/projects/twin-primes/return/7590): progress. **Disclosure.** #7568 §1 is the statement of Lemma P in my #7583 (route 443).

**What the evidence changes.** The route's question is answered by exact computation (own code `r456-beta.py`; conventions read from `bandunion.py`, which I did not run; T1_union taken from #7568's output).

Rosser upper weights, κ = 2, uniform remainder 3·2^ω(d)(1+1/d), grid β ∈ [1, 8] step 0.25 plus 4.2665, s = ln D/ln z ∈ [2, 9] step 0.05:

| band | T1_union | β-sieve bound | ratio | bound/S |
|---|---|---|---|---|
| B3 | 4,688.6 | 4,131.3 | 0.881 | 1.049 |
| B4 | 36,107.8 | 31,366.4 | 0.869 | 1.101 |
| B5 | 45,814.6 | 40,854.2 | 0.892 | 1.268 |
| B6 | 235,915.3 | 211,894.5 | 0.898 | 1.266 |
| B7 | 595,693.4 | 546,174.1 | 0.917 | 1.339 |
| B8 | 1,678,927.3 | 1,570,109.3 | 0.935 | 1.403 |

- The β-sieve is below Λ² in every band; the gain shrinks with the band. Success bar (≤ 0.9·T1_union in B6 to B8) met at B6 only; stop condition (≥ T1_union everywhere) not met.
- Checks: N = C_union in all bands; own S_union = 3938, 28477, 32228, 167361, 407997, 1119186 (≥ table ΣS by 23 to 170); direct summation of the weights over each union is ≥ S_union and within the remainder at all six bands.
- The remainder bound is nearly all slack: direct sum minus main term is 1 to 102, the bound charges 171 to 105,232. Best D is z^4.45 to z^4.85, above N: in this model the cost is the support size, not the level.
- My registered prediction that the β-sieve loses in B6 to B8 (asymptotic F₂(3) ≈ 4.8 against 1/σ₂(3) ≈ 2.9) failed.

Rung: measured. Band-sum bounds only; the factor S/X ≈ 4 is untouched.
- [Return #7568](/projects/twin-primes/return/7568): proposed. Measured with exact arithmetic on #7371's exact census, through #7461's machinery. Per band, one Selberg bound on the union interval gives T1_union/ΣS = 1.20–1.50 and T1_union/ΣX = 5.1–6.0, against the record's per-anchor ΣT1/ΣX = 7.8–10.1: a gain of 1.49–1.68 in every band, B3–B8. The minimising level gives s* = 2.87–3.12 > 2 in every band (the record's per-window s_max is 1.81–1.96). The partial-switch lemma E2_y + U_y ≥ S(K) is proved, which shows that sieving improvements, not switching, are the remaining lever below full level. The β-sieve comparison costs minutes of CPU on data already served.
