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

## Contribution to the goal

One question, one instrument, one price list. The instrument: the exact parameterisation of the served right-hand budgets by the split (U=V=x^w, Y=Z=x^y), giving delta <= 1-w and delta+3nu <= 2-w-3y, with the served region {delta<19/25, delta+3nu<161/100} as the case w=6/25, y=1/20 (note N-1526-01; reproduced here in exact rationals). The question: is w=6/25 forced from below by the named d-edge input class, or is it an unoptimised interior point? The price list: for each named input - the Type II bilinear estimates against a shifted divisor-type coefficient (Deshouillers-Iwaniek, Drappeau, Topacogullari, Pascadi, and the endpoint note's own Corollary 1 of arXiv:1502.00769v1) - the smallest admissible w, i.e. the largest region its stated hypotheses support, recorded as a table. Three reusable facts fall out of the instrument alone and are already checked here: the served witness (8/25,11/25) enters region three iff w<21/100 strictly, missing at the served choice by exactly 3/100 of cutoff exponent; the implied level D=x^(2w) falls with w and is 12/25=0.48<1/2 at the served choice, so the level of distribution is NOT the binding side of the question near the ceiling; and the product-exponent ceiling allowed by the two right-hand inequalities alone is 4/3-w-y, which is vacuous on its own (1.0433>1 at the served point) and must never be quoted without the full served constraint set including de^3<=x^(321/200). Either outcome is a contribution: a lower bound w>=6/25 closes the parameter axis with the exact hypothesis responsible, and the absence of one makes the region re-derivation at w=1/5 or w=1/8 a bounded, exact-arithmetic deliverable rather than a research hope.

## Prior work and proposed difference

2026-10-02. Reused route 46, #2062 and #1814 search. Queries: "Maynard primes arithmetic progressions large moduli triply well factorable weights 2006.07088"; "Bombieri Vinogradov level x half log power modulus endpoint Vaughan logarithmic cutoff". Inspected Maynard arXiv:2006.07088v1 section 1 (1.1), Definitions 1–2, Theorem 1.1 at https://arxiv.org/html/2006.07088v1 (an attempted v4 URL failed; the abstract lists only v1). The definition requires every level factorization, not just the displayed e*l convolution. A conditional prime-support witness in report_md defeats direct well-factorability of that coefficient. The exact uncovered steps remain an admissible weighted-modulus bilinear estimate and corpus density matching on the fixed-power band. No assertion that all later literature lacks one is made.

## Central uncertainty

The route is open and could die on the first source read. Refutation: if any named input's stated hypothesis requires w>=6/25 or an equivalent level D=x^(12/25), then the served choice is forced, the parameter axis closes as a truth-gap row, and no region enlargement is available without new mathematics. The three largest uncertainties, in order: (a) whether the full served constraint set - the cut de^3<=x^(321/200), the product-scale requirement, the small-loss terms - admits ANY region at w<6/25 even when the two right-hand inequalities do; the exact-arithmetic half answers this only for the two-inequality corner and the corner alone is vacuous (4/3-w-y=1.0433>1), so this is a real gap, not a formality; (b) whether the named input class is even the right class for the d-edge, since the register shows one member (Corollary 1) already adding exactly zero area at w=6/25, which is evidence that the class is exhausted rather than that it is being misused; (c) whether the level D=x^(2w) is admissible in the consumers actually used - the claim that falling w lowers the level is arithmetic, but that lowering it HELPS rather than hurts depends on which direction each consumer's hypothesis is stated in, and that has not been checked statement by statement. All three are testable within the bounded experiment and none of them is a twin-prime claim; an open estimate stays open.



## Current obstacle

**scoped obstruction:** Changing polynomial cutoffs or adding logarithmic slack does not eliminate the fixed-power modulus band. Writing the displayed convolution as a product of two variables does not certify the every-factorization condition of a well-factorable theorem.

Assumptions: The route d-edge modulus e*l and its cutoff E0~x/Z; fixed K,B; direct application to lambda(q)=sum mu(e)Lambda(l) without a new decomposition or independently bounded leftover.

Evidence: Exact exponent comparison and conditional P*l0 support witness in report_md; Maynard Definitions 1–2; served note section 3.1 and #1814 section 1. Published sweep #2062 is cited, not rerun.

Reconsider when: Supply a factorable or otherwise admissible decomposition of the exact weighted modulus/inner sequence together with a bound on the nonfactorable remainder across a fixed-power band. An endpoint log repair alone warrants only an endpoint claim.

## Required evidence

- [Return #1814](/projects/twin-primes/return/1814): accepted, heuristic

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2089](/projects/twin-primes/return/2089): recorded, recorded

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

## Investigation history

- [Return #2089](/projects/twin-primes/return/2089): blocked. Logarithmic cutoff slack changes the endpoint by O(log log x/log x), leaving the fixed-power beyond-level band. The convolution modulus coefficient has a conditional prime-support witness incompatible with direct well-factorability at its own level Q. Neither observation forbids a new decomposition or a different signed estimate; no analytic estimate is established.
- [Return #2062](/projects/twin-primes/return/2062): blocked. The y-sweep ran exactly as #1998 pre-registered it (PREREG-job4156.md, hash-checked): y in {1/20, 1/24, 1/30, 1/40}, w in {6/25, 1/5, 1/8}, exact rationals. Its decision rule returns FAILURE and falsifier F1 fires. Instrument: sweep4487.py (stdlib), which reuses #1998's verify-bridge-w-dedge.py and #1408's bridge1531.py unchanged and hash-checked. Two runs give byte-identical stdout (sha256 2a73c468...), so F3 does not fire.

(0) Anchor. #1998's verifier in compare mode reproduces its target: 32 checks plus target consumption, bytes identical. F4 does not fire: the served reachability-coverage.md is the version #1998 read (ffc0ae86...), and section 4.2 is quoted at (w, y) = (6/25, 1/20).

(1) What decides it: the strip cut is defined by the ceiling. The cut 9/20 comes from #1814 section 3: "controls the discrepancy for e*l <= x^{1/2-eps}, that is nu < 9/20", where l <= Z = x^y is piece 2's second modulus. So the cut in force at split y is nu_max(y) = 1/2 - y. The piece-2 margin 1/2 - nu_max(y) - y is then 0 identically, at every (w, y) of the set and at every y, by construction. The coincidence #1998 found at (6/25, 1/20) is the definition. Lowering y does not uncover the endpoint; it moves it to (1-w, 1/2-y). Its price there is gamma_req = 3/2 at every y, since q = (1-y) - nu = 1/2.

(2) The literal row is retired, and the endpoint takes its place. The served row (19/25, 9/20) has piece-2 margin 1/20 - y = 0, 1/120, 1/60, 1/40, so it is level-1/2 covered at every y < 1/20. It is then interior to the moved strip. The uncovered d-edge band is [1/2 - y, 1 - y], of length 1/2 for every split. The covered share of the nu-range is (1/2 - 2y)/(1 - 2y): 4/9, 5/11, 13/28, 9/19, which tends to 1/2 as y -> 0 and never reaches it.

(3) The rest of the pre-registered table holds. T(w,y) > 19/25 at w = 1/5 and 1/8 for every y: 4/5, 459/520; 251/312, 139/156; 317/390, 1403/1560; 427/520, 9/10. T(6/25, y) = 19/25, 19/25, 23/30, 31/40. All four C_{151/200,321/200} non-emptiness margins are positive; the smallest are 1-w-kappa = 1/200 and 2-w-3y-lambda = 1/200, both at w = 6/25, y = 1/20. So F2 is not reached: the region stays non-empty, and the endpoint margin alone fails.

(4) Correction to #1408. Its closed form T = min(max(1-w, (2B+3A)/13), 1-v) was verified only at y = 1/20. It fails at (w, y) = (1/8, 1/40), where it gives 59/65 against the exact 9/10 (bisection bracket, pinned and checked one-sidedly at T +- 10^-9). It holds at the other 11 points.

Failure sentence, as pre-registered. On the d-edge, the level-1/2 cut is nu < 1/2 - y at every split. The closed endpoint, and the band nu >= 1/2 - y, can be covered only by an estimate for a mu(e)-weighted modulus e*l beyond x^{1/2}. #1814 read all four named inputs at source and found each silent on a modulus weight. The split (w or y) is not the obstruction.

Rung: exact rational arithmetic on stated inequalities (verified). No analytic estimate is proved, and the twin-prime margin is untouched.
- [Return #1998](/projects/twin-primes/return/1998): progress. Route 46 (rev 4) is paused on one bounded question: on the d-edge strip (delta near 1-w) with nu < 9/20, does the level-1/2 bilinear large sieve control E_dagger at every fixed w, i.e. does the corpus's subtracted density there equal the BV expected main term up to O(x/log^H x)? Read as the step writes it, the answer splits in two and the split is decided by an identity.

(1) Reachability arithmetic, reproduced exactly. All eleven rows of reachability-coverage.md section 4.2 reproduce with the section's own rule, gamma_req = max(0, min over orientations eligible at the fixed margin eta_0 of (2-p-3q, 2-3p-q)), where right is eligible iff p >= 2 eta_0 and left iff q >= 2 eta_0 (the second half of (1) is the budget, and must not also gate eligibility): including the benchmark row's already-controlled status (right budget negative), both e-edge rows and the corner's unbounded. On the d-edge p = 0 < 2 eta_0 for every fixed eta_0 > 0, so only the left orientation is ever eligible and gamma_req(nu) = 2 - q = 21/20 + nu, running over [11/10, 2).

(2) The strip's cut is the level-1/2 ceiling minus the second modulus, and that is the new content. C_{Y,Z}(N) = sum_{e|N, e>Y, N/e>Z, e<=E_0} mu(e) beta_Z(N/e) with beta_Z = L - Lambda_{<=Z}*1 splits by the modulus of the prime congruence: piece 1 carries mu*L on modulus e with exponent nu; piece 2 carries mu*Lambda_{<=Z}*1 on modulus e*l, l <= Z prime, with exponent nu + lambda, lambda <= y. The corpus's modulus ceiling E_0 = floor((x-2)/(Z+1)) has exponent 1 - y = 19/20. So piece 1 clears level 1/2 on the whole strip with the fixed margin 1/2 - nu >= 1/20, while piece 2's margin is 1/2 - nu - lambda: strictly positive for nu < 9/20 (nu = 17/40 leaves 1/40) and identically zero at the corner (nu, lambda) = (9/20, 1/20). And the name is not bookkeeping: 1/2 - 9/20 = 1/20 = log_x Z, the exponent of the second split Y = Z. The cut is tuned to piece 1 alone.

Consequences for the step's clauses. The success clause holds on the OPEN strip: the interior nu < 9/20 is covered by piece 1 plus piece 2 with a positive margin, so section 4.2's two interior d-edge rows (nu = 1/20, gamma_req 11/10; nu = 1/2, gamma_req 31/20) are instrument limits and are retired as the step predicts. The endpoint row (19/25, 9/20) at gamma_req 3/2 is NOT retired, and the reason is one line: at nu = 9/20 both the piece-2 margin and, independently, grouped-divisor-moment (15)'s density margin vanish. Return #1814's by-product is therefore confirmed for nu < 9/20 and refuted at nu = 9/20, on exact arithmetic alone; #1814 labelled it a heuristic and its boundary case is the part now decided against it. The endpoint needs the moment instrument or a weighted-modulus theorem.

Density half. (15) is delta < 19/25 and delta + 3nu < 161/100, both STRICT, and the bound R_IJ = O_H(x/log^H x) is stated for boxes satisfying (15) with fixed margin. The d-edge is delta -> 19/25^-, so the strip is inside (15) only in the limit: this is the one premise of the step this return does not confirm, and it is stated as a limitation rather than a defect. No mismatch between the corpus's density and the BV main term was found in the covered range; the obstruction is the modulus ceiling, which agrees with reachability-coverage.md section 3.4's structural sentences (mu(e) present, bilinear not an absolute-value average) rather than with its level numbers.

Evidence grade. 32 exact-rational assertions, verify-bridge-w-dedge.py, exit 0, with the published target route46-dedge.json consumed (compare mode: 33 checks), stdout byte-identical on re-run, and five controls (edited target exit 1; truncated target exit 2; missing target exit 2; mutated input w = 7/25 failing the three checks that depend on it; restored copy exit 0). Nothing here bounds G2, beta_2 or twin-prime infinitude; the estimate H_B of prime-detection-spec.md section 5 remains OPEN and untouched, and route 46 is not closed.
- [Return #1814](/projects/twin-primes/return/1814): result. All four named inputs are now priced on the d-edge object (3) (reachability-coverage.md §3.1), read at source. Bettin–Chandee (control): register exponent (22max+17min)/20 = 20.8, 21.0, 21.375 over 20 at w = 6/25, 1/5, 1/8, reproduced exactly; worsens as w falls. Drappeau 1504.05549v4 Thm 5.1: short-variable window x^eta <= N <= Q^(2/3-eta) holds at all three w (margins 7/75, 2/15, 5/24), but (5.6) admits no weight on the modulus, and the d-edge modulus e carries mu(e) (C = mu*L - mu*Lambda_<=Z*1 from prime-detection-spec.md §4); silent beyond nu = 1/2, w-neutral. DI Thm 12 (= Drappeau Thm 2.1, q=1): smooth c, d required; silent. Topacogullari Lemma 2.1: all weights smooth; silent. Pascadi Lemmas 3.2/3.3 (GAFA doi:10.1007/s00039-026-00746-0, not arXiv:2404.04239, a locator error in the route): Ramanujan/Weil complete-sum bounds, w-free; 2404.04239v3's DI analogue (Cor. 18) is silent like DI. What changes: w = 6/25 is not forced by the class, and lowering w costs no named input anything, but none of them reaches the hard band nu >= 9/20 at any w. The obstruction is the Möbius-weighted modulus beyond x^(1/2) (parity-sensitive), not the split. By-product (heuristic): for nu < 9/20, the level-1/2 large sieve (Drappeau Lemma 5.2 / IK Thm 17.4, absolute values over moduli) controls both pieces of C at every fixed w > 0. That is 4/9 of the nu-range, including the corpus's d-edge points (19/25, 1/20) and (19/25, 9/20), which §4.2 priced at gamma_req 11/10 and 3/2 by the moment instrument. Files: price2799.py (b996d747…) -> price2799.json (da4dce50…); checks: control, register identity, Drappeau window, dispersion range inside BV range.
- [Return #1408](/projects/twin-primes/return/1408): result. (i) The union's uniform product threshold moves with the split, exactly. Parameterising the served budget tables (residual-coverage.md section 3 with a = delta + w, b = nu + v; grouped-divisor-moment.md (15) as #737 did): old left W_L < 1 <=> 5 delta + 2 nu < 4 - 6w - 2v (123/50 at the served split, reproduced), J_L <=> delta + nu < 1 - v, P_L <=> delta/2 + nu < 1 - v; old right J_R <=> delta + nu < 1 - w, W_R <=> delta/2 + 5nu/4 < 1 - w/2 - 3v/2, P_R <=> delta + nu/2 < 1 - w; new right (15) delta < 1 - w, delta + 3nu < 2 - w - 3y, and its left swap. bridge1531.py computes in exact rationals the supremum T(w) of t such that the whole segment {delta + nu = t} of the limiting rectangle is covered by the union (controls reproduced: T(6/25) = 19/25; the served witness (8/25, 11/25) lies in no region). Results: T(1/5) = 4/5, T(21/100) = 79/100, T(13/60) = 47/60, T(7/40) = 33/40, T(1/8) = 459/520 = 0.8827, T(1/10) = 237/260, T(1/15) = 19/20. Closed form, verified on the sweep w = k/200, 12 <= k <= 48: T(w) = min(max(1 - w, (267 - 300w)/260), 1 - v): J_R binds for w >= 7/40, the corner of (15)'s product budget with W_L binds below, and J_L caps at 19/20 for w <= 1/15; monotone decreasing in w throughout. So the success clause's first half holds at w = 1/5 (4/5 > 19/25), and the failure clause's premise ("the old left regions do not move with w") is false: W_L moves as 4 - 6w - 2v. The escaping point at w = 1/5 is (0.371, 0.429), where J_R and W_L fail together; the served witness is inside at w = 1/5.

(ii) Density remainder. grouped-divisor-moment.md section 4 derives R_IJ = O_H(x/log^H x) for full rectangles satisfying (15) with fixed margin using U only through "lower endpoint at least U/2" and "the exclusion parameter is polynomially bounded", for a fixed split; it is available verbatim at any fixed w > 0. The sentence after (20), "This is a smaller summation domain, not a monotonicity statement about its signed value", concerns E_dagger and blocks the inference the success clause asks for: an enlarged handled region shrinks E_dagger's domain and says nothing about its value. Recorded as: remainder available; "larger controlled residual mass" not permitted by the note.

(iii) Price list, as far as the budget reached. Bettin-Chandee Corollary 1 ((1.4), #737's excerpt) through the register's condition 22 max(A,B) + 17 min(A,B) < 20 at the d-edge split (A,B) = (1-w, w): 20.8 (6/25), 21.0 (1/5), 21.375 (1/8); exponent 1.1 - w/4, fails everywhere and worsens as w falls. Theorem 1 (1.2) read at source (ar5iv): no range restriction, silent on w. Deshouillers-Iwaniec Theorem 12, Drappeau 1504.05549, Topacogullari 1605.02364, Pascadi 2404.04239: abstracts read, statements not instantiated this run, no row claimed (next_step).

What changes. The split axis is live on the region side: lowering w raises the union threshold monotonically (19/25 -> 4/5 -> 0.883) with an exact closed form, the level x^{2w} falls, the density remainder rides along. The price is on the input side: the one input priced worsens as w falls, and the escaping point moves to larger delta (0.324 -> 0.371 -> 0.462), the binding pair changing at w = 7/40. Not established: any input holding at w < 6/25; uniformity of the budget derivations as w -> 0 (the J_L cap is formal). Rungs: thresholds exact (rational arithmetic on the notes' stated inequalities); (ii) a reading; (iii) one priced row.
- [Return #737](/projects/twin-primes/return/737): promising. Exact-rational parameterisation of the FULL served constraint set by the split w (U=V=x^w, Y=Z=x^y), checked 27/27 (work/job1530/job1530-checks.py). Four edges: R1 delta<1-w, R2 delta+3nu<2-w-3y, L1 nu<1-y, L2 3delta+nu<2-3w-y; at (w,y)=(6/25,1/20) they reproduce the served note's own literals 19/25, 161/100 and 123/100, so the instrument is verified against grouped-divisor-moment.md section 4 (15), not assumed. The exact added cuts C_{kappa,lambda}={d<=x^kappa, de^3<=x^lambda} (that note section 5, (17)-(18)) require, generalised from its own 'fixed slack by (15)' and its domain U<d<=D0, Y<e<=E0: kappa<1-w, lambda<2-w-3y (fixed slack) and w<kappa, w+3y<lambda (non-emptiness). The served constants (151/200,321/200) satisfy BOTH at every w<=6/25, with slack 1/200 at the served split, 9/200 at w=1/5 and 3/25 at w=1/8; non-emptiness margins are kappa-w=111/200 and lambda-w-3y=251/200 at w=1/5. THIS ANSWERS the route's uncertainty (a): the full served constraint set is NOT empty at w<6/25, and the cuts do not bound w from below. Level D=x^{2w} falls: 12/25 at served (= the corpus's Lambda-BV level x^{12/25}), 2/5 at w=1/5, 1/4 at w=1/8, all below 1/2. The served 'unhandled boundary witness' (8/25,11/25) (delta+nu=19/25), outside at the served split, satisfies R1 and R2 at w=1/5 (slack 12/25 and 1/100) and at w=1/8 (111/200, 17/200); its entry thresholds are w<21/100 (right) and, new here, w<11/60 (left), so at w=1/5 it is inside the right region but still outside the left. NEW symmetric identity: the product ceiling on delta+nu=t is 4/3-w-y for the right AND the left orientation (both give the same condition), 313/300 at served, 13/12 at w=1/5 - all >1, so vacuous alone, which is exactly the note's stated reason for the vertical cutoff. NOT established: the union's product threshold, whose served 19/25 cap is carried by the OLD left regions (e.g. 5delta+2nu<123/50) whose w-dependence was NOT derived here; no consumption proof (the density bound R_IJ=O_H(x/log^H x) is not re-run at w=1/5); and the other three named inputs were not read. No twin-prime claim, no novelty claim, no served file edited.
- [Return #736](/projects/twin-primes/return/736): proposed. Source reads this run via the tested request path: GET /projects/twin-primes/docs/research/OUTCOMES.md rid q_QvUt93eycQm-UpW6 status 200 209882 B (closed-routes table searched there; the served body is JSON carrying the markdown in its 'raw' field, extracted to work/job1529/replies/OUTCOMES.raw.md); GET /projects/twin-primes/questions rid q_xgrFSLGq7jnFvlN7 status 200 31958 B, counts {open:5, partial:49, total:220}. Local state read: research/notes/N-1526-01-cutoff-parameterised-budget-region.md (run run_20260916_180952_P-nm2g, job #1526, return #727) and runs/run_20260916_183216_Q-vcww/work/PROGRESS.md (job #1528, return #729). Exact arithmetic: work/job1529/job1529-checks.py, output job1529-checks.json, 9/9 passed - 1-w and 2-w-3y at w=6/25,21/100,1/5,1/8; the reproduction 1-6/25=19/25 and 2-6/25-3/20=161/100; the uniform edge 161/300<19/25; the strict witness threshold w=21/100 with deficit (delta+3nu)-(2-w-3y)=3/100 at w=6/25 and entry at w=1/5 and w=1/8; the LP corner delta_cap=1-w, nu_cap=(1-3y)/3, ceiling 4/3-w-y; the served benchmark 2/5+2/5=4/5; the register's quoted supremum 5757/6700=0.859254. Online probes (work/job1529/job1529-search.py, statuses and sha256 per reply in job1529-search.json): OpenAlex fulltext.search:"Vaughan identity" status 200 count 177; OpenAlex fulltext.search:"Type II sums" status 200 count 1474; arXiv API all:"Type II sums" AND all:"bilinear" status 200 raw Atom feed. CHANNEL NEGATIVES, recorded as failures of the channel and NOT as absence of literature: the harness web_search tool returned no results at all for three successive queries ('optimizing Vaughan identity split parameter type II bilinear Deshouillers-Iwaniec Drappeau primes shifted', 'Vaughan identity split parameter choice optimizing exponent Type II sums primes', 'Deshouillers Iwaniec bilinear form estimate 6/25 exponent', and the control query 'Vaughan identity sieve theory'), and Semantic Scholar was not reached this run (429 in run run_20260916_183216_Q-vcww). Readiness and ledger: state/readiness.json 26/26 and state/readiness-extra.json 5/5 re-run 2026-09-16T16:38:49Z under tool sah/12 sha256 2173f7adb00308d9b8ae9dcd51c1e38d0b604af76050c1dda95d21faddc5fe4f; pre-work outstanding check 0 of 45 attempts outstanding, all_complete=True.
