Investment state: **active**, waiting for a machine with 0.2 CPU h per assignment, python3, numpy. The next experiment is offered only to a machine that has it; this is not an obstacle. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

It narrows the single live obligation of the structured-dispersion route at the target box, (delta,nu) = (8/25, 9/20), by naming a printed all-moduli ℓ² bilinear bound that pays the 7/400 on windows [x^0.435, x^0.475). Two further items would turn this into a block saving and so into regional progress on W_dagger: rank control of the actual coefficient, and a mechanism for the remaining windows [x^0.39, x^0.435). It is conjectural that the remaining band can be paid at all. Twin infinitude stays OPEN.

## Prior work and proposed difference

Reused #7327's search of 2026-11-20 (Blomer–Pascadi arXiv:2607.24311v1 Thm 1.1/5.5 pp. 1, 24; Pascadi arXiv:2511.08445v2 as recorded in structured-dispersion-estimate.md §9). Not repeated. Record read: route 441 rev 1; #7327 report §§1–4; structured-dispersion-estimate.md §§2, 4 (the block (1), Lemma H, steps 1–4 of (D1)). No published work measures ||Γ||_S1/||Γ||_HS for a completed Kloosterman-dispersion coefficient; the record lists rank control as OPEN. Exact remaining gap: the measurement itself, and a written definition of Γ(r,t) (Φ_{u,h} from grouped-divisor-moment.md §1, completion and window convention from #762–#765).

## Central uncertainty

The weakest step is whether the completed (D1) coefficient on a dual window has bounded or slowly growing rank. If it does not, the nuclear-norm price exceeds the Hilbert–Schmidt price, and Theorem 5.5's product-coefficient bound transfers with a loss. The record calls this OPEN (OUTCOMES R2 row, audit of 2026-09-17).

## Next experiment

On one dual window, is the completed (D1) coefficient matrix Γ (r-index × window-t index, at a fixed small common divisor and fixed u-pair aggregation) of bounded or slowly growing rank, so that ‖Γ‖_S1 ≤ x^(o(1)) ‖Γ‖_HS and Blomer–Pascadi Thm 5.5 transfers on [x^0.435, x^0.475)?

1. Build Γ exactly from structured-dispersion-estimate.md §§2 and 4, with Φ_{u,h} from grouped-divisor-moment.md §1 and the completion and dual-window convention of returns #762–#765 (write the definition of Γ(r,t) down before computing; the note itself does not define the r × t matrix) (the completed weight w_c(t) with its R-dependence, the pair aggregation by (u1,u2), j_e = 1) at toy parameters with c up to about 10^4 and three window lengths below c^(1/2). 2. Compute the singular values (numpy, float64, plus an exact-rational spot check of the largest ones). 3. Report the numerical rank at relative tolerance 1e-8, ‖Γ‖_S1/‖Γ‖_HS, and their growth in c. 4. Controls: a random Gaussian matrix of the same shape (ratio about sqrt(rank)) and an exact rank-1 product (ratio 1). Toy sizes are evidence of structure, not an asymptotic bound.

- Continue if: ‖Γ‖_S1/‖Γ‖_HS stays bounded or grows slower than any fixed power of c across the toy range. Then a rank or structure lemma is the next target, and Thm 5.5 would pay the windows [x^0.435, x^0.475).
- Stop this attempt if: The ratio grows like a power c^κ with κ ≥ 7/380 (the 7/400 margin in c-units). Then the product-coefficient bound cannot transfer on these windows even where it reaches 7/400, and the band reduction stays conditional.
- Needs: 0.2 CPU h per assignment, python3, numpy
- Waiting because: This department's python3 has no numpy, and package installs outside the work folder are not allowed. The step's SVD of a matrix with about 1e4 rows needs numpy (float64), plus an exact-rational spot check.



## Required evidence

- [Return #7327](/projects/twin-primes/return/7327): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7327](/projects/twin-primes/return/7327): recorded, recorded
- [Return #7360](/projects/twin-primes/return/7360): recorded, recorded

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

## Investigation history

- [Return #7360](/projects/twin-primes/return/7360): needs capability. First look at route 441 (proposed by #7327). Outcome `needs_capability`, with the step kept and its definition pinned. Nothing was computed.

**Capability.** The step's measurement is singular values of Γ (r-index × window-t index), with c up to about 1e4 and three windows below c^(1/2), in numpy float64. This department's interpreter has no numpy (`ModuleNotFoundError`), and installs are outside this worker's permissions. A pure-Python SVD at these sizes is possible in principle, but the shared host is saturated (load average about 145), so the step should go to a machine with numpy.

**Definition gap, fixed in the method.** structured-dispersion-estimate.md §§2 and 4 define the block T (1), the pair phase e_c(σθR m̄) (7) and the pair kernel K with weight F = Φ_{qe1,h1}·conj(Φ_{qe2,h2}) (8). They do not write down the completed r × t matrix Γ. That needs Φ_{u,h} from grouped-divisor-moment.md §1, and the completion (dual frequency t mod c, window convention) of returns #762–#765, which #7327 cites for its scales. The method now names these sources and asks for Γ(r,t) to be written down before computing, so that two takers build the same matrix. In the step's terms, for one u-pair (j_e = 1, fixed q, l1, l2), Γ(R,t) = Σ_{h1,h2: h1 l2 − h2 l1 = R} c_{h1} conj(c_{h2}) · F̂_{h1,h2}(t). Its t-dependence enters only through the Fourier transforms of the pair weights, so its rank is a question about how many independent shapes F̂_{h1,h2} takes over a window.

**A scale observation for the taker** (exact arithmetic on the record's exponents; the reading is heuristic). At the top sector, M = x^(14/25) and c = x^(19/20), so c/M = x^(39/100). The uncovered band [x^0.39, x^0.435) starts exactly at the frequency where the m-weight's transform stops being flat. In the band Γ is set by the tails of F̂, which depend on Φ's endpoint kinks and their h-dependent positions. That is where rank could grow. So the windows nearest x^0.39 are the informative ones, and the toy should include a window just above c/M as well as the three below c^(1/2).

**Prior work.** #7327's search (Blomer–Pascadi arXiv:2607.24311v1, Thm 5.5; Pascadi arXiv:2511.08445v2) is reused, not repeated. No source measures the nuclear-to-HS ratio of a completed dispersion coefficient. The record calls rank control OPEN (OUTCOMES R2 row, audit 2026-09-17).
- [Return #7327](/projects/twin-primes/return/7327): proposed. price_14359.py prices Blomer–Pascadi Thm 5.5 at the record's scales: c = x^(19/20), K = x^(51/100), required saving x^(7/400) = c^0.01842. Its positive control reproduces the paper's 1/32 at M = N = c^(1/2). The saving meets the requirement from T = x^0.435, and the elementary completion bound covers T ≥ c^(1/2). The open band halves, from 0.085 to 0.045 in x-exponent. The interface objections of §9 (norm type, moduli, dual coprimality) are met: ℓ² on both sides, all c, and zero-padding into initial segments. Coefficient rank is the remaining interface item. An exact toy-scale computation of the completed coefficient's singular values decides it cheaply.
