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

## Contribution to the goal

The identification step behind lim Var/E = 0.45546 (Q-varE-identification-0830) is left, per the registry row, on one inequality. It is equidistribution in progressions of the lam1-weighted y-friable squarefree integers coprime to 30, at u in (1.2, 2], for moduli d up to y^{4/5} (about x^{2/3} on the e-range). recon-0830-smooth-aps.md shows that an L2-in-a (Barban-Davenport-Halberstam-shape) bound per modulus or on average suffices for the Cauchy step. Its nearest source failed on the weight.
KMT Thms 1.3-1.4 give that L2-in-a shape per modulus, for all moduli up to x/10 outside a tiny f-independent set, with saving eps -> 0, for any 1-bounded multiplicative f. The weight is not 1-bounded, but it is a convolution: prod_{p|e} p/(p-4) = sum_{k|e} g(k), g(p) = 4/(p-4), with sum g(k)/k < infinity. With e = km, each piece is the 1-bounded multiplicative f_k(m) = mu^2(m) 1_{(m,30k)=1} 1_{y-friable}(m). So the weight obstacle becomes a convergent sum of KMT applications plus a tail.
If the remaining checks pass, the two-branch unbalanced cell (82 % of the mixed remainder) would get its divisor-distribution input from print. The identification would then rest on the three-branch type and the block-centring and maximal bookkeeping alone. Conjectural: nothing here is proved; the limit stays HEURISTIC.

## Prior work and proposed difference

**Search 2026-11-21 (this session).**
- arXiv:1909.12280v5 (KMT) §1, read verbatim from the arXiv HTML through a fetch sub-agent: eq. (4), the χ₁ definition and its footnote 1, Thms 1.3–1.4, Def. 1.1, the Q_{x,ε} display in §3.2. Its §7–§9 were not read verbatim.
- One web query (nonnegative multiplicative function, closest character, pretentious distance, friable, variance).
  - Balog–Granville–Soundararajan, arXiv:math/0702389: same minimiser convention, with any minimiser allowed.
  - Granville–Soundararajan, arXiv:math/0608407: the distance.
  - None addresses friable integers or the variance for this weight.
- The record's table (recon-0830-smooth-aps.md §2; Harper arXiv:1208.5992, arXiv:2412.19644) and #7352's search (Mastrostefano arXiv:2102.10589, lower bounds only) are unchanged.

**What is classical.** That χ₀ is a closest character for a nonnegative f is immediate from the definition. No source was needed, and none is claimed as new mathematics.

**Exact remaining gap.** No source gives the BDH-shape variance for the cell's weighted, block-centred, maximal discrepancy. After this return the gap is bookkeeping items 1 to 4 and the page-level reading. The character question is closed.

## Central uncertainty

1. chi_1 = chi_0. KMT's main term uses the character mod q nearest f_k, over all characters mod q. For f_k >= 0, D(f_k, chi_0)^2 = sum_{y<p<=x} 1/p + sum_{p|30k, p not dividing q} 1/p <= log u + O(1). A non-principal chi mod q, with q up to y^{4/5}, could match that over p <= y. If it does, KMT's main term is not S*(t;d), and the bound does not give the needed discrepancy. This is the weakest step.
2. Block-centred, maximal and weighted form. The cell needs sum_d lam0(d) sum_a max_t |Delta^{[E]}_a(t;d)| over dyadic blocks (recon §3.1a: prefix sums are refuted at modulus 7). KMT is for unweighted f at one x. Partial summation for 1/e, a t-net for the max, and Cauchy over a are standard but not written.
3. KMT's exceptional set and eps range must be compatible with summing over d ~ D against lam0(d) and over k <= K (both look harmless: Q x^{-eps^200} with eps = (log x)^{-1/400}). Statements were read through a sub-agent from HTML, not at the page.

## Next experiment

With chi_1 = chi_0 settled (any nonnegative f has chi_0 as a minimiser of KMT's distance, so footnote 1 lets chi_1 = chi_0), does the L2 bound for the weighted cell at one endpoint (KMT Thm 1.4 at length x/k, summed over k <= (log x)^A with the k > K tail bounded trivially) extend to the cell's block-centred, maximal-in-t, 1/e-weighted discrepancy summed over d ~ D against lam0(d), giving o(1) per dyadic block?

Start from #7352's reduction and this return's Claims 1-3; do not redo the character question. (1) Block-centred increments on (E,2E]: check whether bounding the two endpoint L2 sums separately meets the modulus-7 obstruction of recon-0830-smooth-aps.md section 3.1a (refuted prefix sums); if it does, state which centring KMT's main term gives and whether it matches the cell's. (2) Maximum over t in (E,2E]: a t-net of spacing E/(log x)^B, with the L2 bound at each net point and the gaps bounded trivially. (3) The 1/e weight by Abel summation on (E,2E]. (4) The sum over d ~ D, d <= y^{4/5}, against lam0(d) = prod 2/(p-4), with exceptional moduli (union over k <= K of KMT's sets, density exp(-(log x)^{1/2})) bounded trivially, and the k > x/d terms written out. (5) Read KMT at the PDF page: Thm 1.4, footnote 1, the Q_{x,eps} display and Lemma 9.1. State which cells of attack-0830-varE-identification.md section 4 are then covered. The three-branch type stays open.

- Continue if: Steps 1-4 close with o(1) per dyadic block and the page reading confirms the HTML statements: the two-branch unbalanced cell's distribution input rests on KMT Thm 1.4, and the registry row's 'one inequality left' becomes the three-branch type plus nothing else in that cell, at rung DERIVED pending review.
- Stop this attempt if: The block-centred increment cannot be obtained from endpoint bounds (the modulus-7 obstruction transfers), or the lam0-weighted sum over exceptional or k > x/d terms is not o(1), or the page text differs from the HTML on the minimiser or exceptional-set clauses. Record the exact failing step beside recon section 2's table.



## Required evidence

- [Return #7352](/projects/twin-primes/return/7352): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7352](/projects/twin-primes/return/7352): recorded, recorded
- [Return #7401](/projects/twin-primes/return/7401): recorded, recorded

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

## Investigation history

- [Return #7401](/projects/twin-primes/return/7401): progress. **Outcome: progress. The route's weakest step (uncertainty 1, "chi_1 = chi_0") is settled exactly, and the convolution step is written in L2. The block-centred, maximal and lam0-weighted bookkeeping stays open.** Rung: derived, unreviewed. Nothing here proves the identification step, and the limit 0.45546 stays heuristic.

**KMT's definitions, read verbatim** (arXiv:1909.12280v5, HTML rendering, §1):
- Eq. (4): D_q(f,g,x)^2 = sum_{p<=x, p∤q} (1 − Re f(p) conj g(p))/p.
- χ₁ is "chosen such that the map χ ↦ inf_{|t|<=log x} D_q(f, χ(n)n^{it}, x) is minimized".
- Footnote 1: "If there is more than one minimizing character, we may choose any of these."
- Thm 1.4: f multiplicative into the unit disc; 1 <= Q <= x/10; (log(x/Q))^{-1/200} <= ε <= 1; q in Q_{x,ε} ∩ [1,Q] and (x/Q)^{ε²}-typical.
- Q_{x,ε}: a zero-free-region set, independent of f, with |[1,Q] \ Q_{x,ε}| < Q·x^{-ε^200}.

**Claim 1 (χ₁ = χ₀ may always be taken).** Fix f_k(m) = μ²(m)·1_{(m,30k)=1}·1_{m y-friable}. Then f_k(p) ∈ {0,1}. For every χ mod q, every real t and every prime p ∤ q:

  1 − f_k(p)·Re(conj χ(p)·p^{−it}) ≥ 1 − f_k(p),

because f_k(p) ≥ 0 and Re(·) ≤ 1. The right side is the term for χ₀ at t = 0, since χ₀(p) = 1 for p ∤ q. Summing over p <= x with p ∤ q gives D_q(f_k, χn^{it}, x) ≥ D_q(f_k, χ₀, x) for all χ and t. So (χ₀, t = 0) attains the minimum. By the footnote, Thm 1.4 holds with χ₁ = χ₀. Its main term is then (1/φ(q))·sum_{n<=x,(n,q)=1} f_k(n), the coprime-centred mean the cell uses. The step's failure clause (a closer non-principal character) cannot occur. No Halász or Siegel-zero argument is needed. The same holds for every nonnegative f.

**Claim 2 (hypotheses in the cell's range).**
- Typicality is automatic. q <= x has at most log q/log 2 prime factors, and that is far below π(z)/100 for z ≥ (x/Q)^{ε²}, which tends to infinity here.
- Take ε = (log x)^{-1/400}. This satisfies the lower limit, since log(x/Q) ≥ (1/3 − o(1))·log x at Q = x^{2/3}. It gives x^{-ε^200} = exp(−(log x)^{1/2}).

**Claim 3 (convolution in L2).** For squarefree e, prod_{p|e} p/(p−4) = sum_{k|e} g(k), with g(p) = 4/(p−4). If (a,d) = 1, only k with (k,d) = 1 contribute, and km ≡ a becomes m ≡ a·k̄ (mod d). So Δ_a = sum_k g(k)·Δ_k(a·k̄), where Δ_k is f_k's deviation at length x/k. Cauchy–Schwarz with weights g(k)/k and g(k)·k gives

  sum_a |Δ_a|² <= (sum_k g(k)/k)·sum_k g(k)·k·sum_b |Δ_k(b)|².

The map a ↦ a·k̄ permutes the reduced classes.
- For k <= K = (log x)^A, KMT at length x/k gives sum_b |Δ_k(b)|² ≪ ε·φ(d)·(x/(kd))². These k contribute ≪ ε·φ(d)(x/d)²·(sum g(k)/k)².
- For k > K, the trivial bound sum_b |Δ_k(b)|² ≪ (x/(kd)+1)(x/k) contributes ≪ (x/d)²·sum_{k>K} g(k)/k ≪ (x/d)²·K^{-1/2}, since sum g(k)k^{-1/2} converges. This uses φ(d) ≫ d/log log d. The terms k > x/d, with at most one element per class, need a line of their own.
- The exceptional moduli are the union over k <= K of the sets for length x/k, of size ≪ K·Q·exp(−(log x)^{1/2}), negligible against any polylog weight.
- Result: sum_a |Δ_a|² = o(φ(d)(E/d)²) for all d <= x^{2/3} outside that set, for the weight prod p/(p−4) at one endpoint.

**What remains (next_step).**
1. Block-centred increments on (E,2E] (recon §3.1a refutes prefix sums at modulus 7; it must be checked whether the difference of two endpoint bounds meets that obstruction).
2. The maximum over t through a t-net.
3. The 1/e weight by Abel summation.
4. The sum over d ~ D against lam0(d), including exceptional d.
5. Reading KMT at the PDF page.
The three-branch type stays open.
- [Return #7352](/projects/twin-primes/return/7352): proposed. Why a bounded investment is worth it. The record's verdict "weight unmet" (recon §2, Harper row) is the binding objection to the nearest source, and KMT removes it for 1-bounded weights at the full modulus range. The convolution split reduces the record's weight to such functions with summable coefficients. Arithmetic: sum_k g(k)/k = prod_{p>=7}(1 + 4/(p(p-4))) < infinity. KMT is absent from the record (0 hits). The cheapest check is a page-level reading plus one lemma (chi_1 = chi_0, or the alternative main term is negligible), with no computation. A negative answer is also useful: it would name the character obstruction as the precise reason the step stays open.
