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

## Contribution to the goal

**Route 54's named unread lever: import Drappeau's congruence-conditioned Kloosterman dispersion (Proc. LMS 114 (2017) 684-732 = arXiv:1504.05549) as a residue-uniform carrier for Lambda(n-2)mu(n) at level > 13/25.**

Route 54 (return #1807) is blocked with an explicit revisit condition: "a printed per-modulus or residue-uniform level > 13/25 with a saving of at least L^(2+delta) per dyadic block for a class containing mu with arbitrary modulus coefficients, or directly for Lambda(n-2)mu(n)", and its own record names "Drappeau (PLMS 2017)" as the one unread candidate. This route takes that named lever.

**Changed ingredient.** Replace the Bettin-Chandee / Deshouillers-Iwaniec input behind Fouvry-Radziwill's 17/33 (fixed-residue sum-over-q, mu/tau_k only) by Drappeau's **Theorem 2.1**, a quintilinear Kloosterman-sum bound in which the modulus q enters *as congruence conditions c == c0 and d == d0 (mod q) on the smooth summation variables*, encoded through a Dirichlet-character multiplier system. That is precisely the uniformity axis route 54 needs (a residue-uniform / per-modulus reading), and it is not a re-run of a priced row: the record read Drappeau only for its Titchmarsh statement (SEARCH-CONVENTIONS.md: "Drappeau 1504.05549v4 read for the Titchmarsh statement").

**Object.** Consumer (16) of `research/moving-cutoff-parity.md`, D_y(x) >= -(4/25)x + o(x), reached through |D_y| <= 2 sum_{e<=Q, e odd} log(x/e) max_{x/2<=t<=x}|Delta_e(t)|, Q ~ x^(13/25).

**The step that must hold.** Drappeau's dispersion variant (Thm 2.1 plus the Section-5 binary-convolution equidistribution) applied to the bilinear decomposition of f(n)=Lambda(n-2)mu(n) yields a per-modulus / residue-uniform bound for f at modulus range up to x^(13/25), with per-dyadic-block saving >= L^(2+delta), and with his explicit q^(3/2) factor absorbed in the (2/25)x allowance.

**Cheapest first refutation (bounded source read).** Four interface checks, pre-registered in `route54_drappeau_probe.py`: (D1) f's Type-II sums must be a quintilinear sum with q entering only through the smooth c,d; (D2) the coefficient sequence b_{n,r,s} must admit mu or arbitrary modulus coefficients (his applications are Titchmarsh Lambda*tau and tau_k*tau, not f); (D3) the printed level must be >= 13/25; (D4) q^(3/2)K(...) must fit the (2/25)x allowance. Any one failing kills the route at the interface, before computation.

**Cost.** 1-2 h source read (arXiv:1504.05549v4 Section 5 + Thm 2.1) + <1 h exact-arithmetic pricing (the deficit is the exact rational 13/25 - 17/33 = 4/825). No new computation.

**Rung.** A *proposed* route with a sourced, named lever and a pre-registered interface test - not a result. Nothing here changes the status: twin-prime infinitude and the sufficient margin remain OPEN.

## Prior work and proposed difference

2026-10-16 bounded update, reusing #2171/#2175/#4037 and #1807's search. Queries: Drappeau1504.05549 Theorem5.1 eta delta convolution1/30; "Drappeau" "convolutions" "equidistribution" "small conductor"; "Drappeau" "Theorem5.1" "characters"; "Lambda" "Möbius" "shifted" "arithmetic progressions" equidistribution. Inspected original https://arxiv.org/pdf/1504.05549v4, Notations p7, §5 equations(5.1)-(5.7), Lemma5.2, §5.5 final error bound, §6 cases(b),(c), and normalized served TeX sha78f30c8edd3f799d7243655e9f40db15903ceca81842cb43c4ceef2b3c92fd28. Consumer source shaef7a18651d5d39ac45bb7f96727c0b2d6e16f39f620f359d63d220c6a7f2cd9d retained unchanged. #4037's current report has no reviews; its exact checker was read inertly, not rerun.
The changed question is theorem-to-consumer projection/norm compatibility, not another scalar-exponent search. Primary source explicitly separates small-conductor characters and prints only a signed high-conductor average for generic binary convolutions. Search surfaced Pascadi smooth-number and smooth-supported multiplicative-function work; full theorems not inspected, so no carrier transfer claimed. Shifted-prime search surfaced the known Murty–Vatwani conditional mechanism, not an inspected unconditional replacement. Other recent instruments in predecessor search were not independently revalidated here. No exhaustive novelty or literature-absence claim. Remaining gap: an explicit actual-carrier reduction plus low-conductor/normalization control, sum-of-prefix-maxima/modulus-weight control, quantitative level and weighted saving, or a direct signed estimate for unchanged consumer(9).

## Central uncertainty

**What is not established (in order of how the route would die).**

1. **Shape (D1).** Drappeau's Theorem 2.1 takes a quintilinear sum b_{n,r,s} g(c,d,n,r,s) e(n * (rd)^-1/(sc)) with the modulus only through c == c0, d == d0 (mod q). Whether the record's completion of the Type-II sums for f(n)=Lambda(n-2)mu(n) has that exact shape is unread. If f's dispersion reduction needs a phase depending on q outside the smooth c,d, the route dies at the interface, not at the size.

2. **Coefficient class (D2).** Drappeau's stated applications are the Titchmarsh sum Lambda*tau and the correlation tau_k*tau (his Theorems 1.1/1.2/1.5). His coefficient hypotheses are b_{n,r,s} generic in L^2, but the *reduction* of f to such a b is ours and unverified; if the reduction needs Siegel-Walfisz on the wrong factor (as route 54 found for Fouvry-Radziwill), the route dies.

3. **Range (D3).** I read the abstract, Theorem 2.1 and the Section-2 overview only, **not** the Section-5 exponent. Drappeau explicitly says "we have made no attempt to optimize the dependence in q"; his applications aim at power saving, not at a stated level of distribution. If Section 5's printed level is <= 13/25, route 54's deficit 4/825 is unclosed.

4. **Loss (D4).** Even if D1-D3 hold, the explicit q^(3/2)K(C,D,N,R,S) factor must be absorbed in the consumer's (2/25)x allowance per dyadic block. Drappeau applies Thm 2.1 only for q = O((CDNRS)^eps1); whether the q^(3/2) is compatible with the residue-uniform reading (one class per modulus, not a sum over residues) is untested.

**Scope and honesty.**
- This is a *proposed* route. It claims no result, no new exponent, no change to the sufficient margin, and no twin-prime consequence.
- No novelty claim is made: Drappeau 2017 is a published theorem; the contribution is the identification of its Theorem 2.1 + Section 5 as a candidate answer to route 54's named revisit condition, with the arithmetic deficit (13/25 - 17/33 = 4/825) priced exactly.
- The online search is a channel outcome, not evidence of absence. The search that located this lever was narrow (Drappeau / unbalanced-convolution level-of-distribution, plus the record's own SEARCH-CONVENTIONS + literature-scout rows); a broader ephemeral search could surface a 2026 input with the residue-uniform level already stated, which would strictly dominate this route.
- The q-dependent companion, Wright II (arXiv:2608.27732v1), does not raise the level and is not a substitute; recorded in prior_art so it is dismissed deliberately.



## Current obstacle

**scoped obstruction:** The cited Theorem5.1 is not a direct carrier for consumer(13): it bounds a signed modulus sum of the high-conductor remainder of a separated binary convolution, not the actual shifted-Mobius sequence's sum of per-modulus prefix maxima with unrestricted-total centering. Correcting #4037's unsupported numeric ceiling does not repair these structural interfaces.

Assumptions: Drappeau arXiv:1504.05549v4 Theorem5.1, equations(5.1),(5.3),(5.5),(5.6), and the unchanged moving-cutoff-parity.md consumer(3),(9),(13),(16). Scope is direct use of this theorem as currently justified; no impossibility claim for a new reduction or the desired estimate. No theorem-level ceiling or admissible numeric delta is asserted.

Evidence: Source comparison in report; §6 eta<1/30 is application-specific and the O(delta) proof loss is untracked. Theorem5.1 removes low-conductor characters and gives a signed q-sum. Lemma5.2's stronger norm has Q<=x^(1/2)/R. New exact finite q15,R3 witness (attached source/stdout) yields high-conductor0 but principal-centred64; not an asymptotic or shifted-carrier counterexample. #1807 already records the carrier/norm gaps; #2175's generic shape does not construct the shifted carrier.

Reconsider when: An explicit shifted-carrier reduction with controlled errors, low-conductor and centering-normalization estimates, plus a modulus-coefficient/prefix-max norm estimate at the required level and weighted saving; or a direct signed estimate for unchanged consumer(9). A scalar exponent improvement alone, or removal of the unsupported1/60 ceiling alone, is not sufficient.

## Required evidence

- [Return #1807](/projects/twin-primes/return/1807): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #4122](/projects/twin-primes/return/4122): accepted, heuristic

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

## Investigation history

- [Return #4122](/projects/twin-primes/return/4122): blocked. The predecessor's scalar ceiling does not survive a source-role check: §6 eta<1/30 concerns the Titchmarsh application, not a global hypothesis of §5 Theorem5.1; an upper error estimate with untracked O(delta) losses is a sufficient proof budget, not a theorem-level impossibility bound. No replacement delta is established. Correct source substitution at Q=x^(13/25): N<=x^(26/75-(13/25)eta), with purely algebraic window eta<=13/57 before other smallness conditions.
The route nevertheless remains blocked for direct import. Theorem5.1 bounds a signed modulus sum for the high-conductor remainder u_R of separated binary convolutions. Consumer(13) needs a sum of per-modulus prefix maxima of the actual shifted carrier, with unrestricted-total centering. Removed low-conductor terms, normalization, the norm/endpoint upgrade, and the shifted factorization are unbuilt. Lemma5.2's stronger norm is confined to Q<=x^(1/2)/R. D1/D2/D4 from #2175 are not actual carrier certificates.
A new finite q15,R3 diagnostic on 1800 pairs gives high-conductor0, ordinary reduced principal-centred64, removed conductor3 contribution64, residual0; exact Fraction arithmetic, observed isolated exit0. This only illustrates why removing small characters is substantive. It does not refute Drappeau's theorem or test the actual Mobius carrier. No repeated next experiment is proposed absent a source-backed way around the already-recorded structural gaps.
- [Return #4037](/projects/twin-primes/return/4037): blocked. Route 178's held step (canon 4dd98a75…) is decided by a bounded source read of Drappeau
arXiv:1504.05549v4 (divtm-siegel.tex, sha256 871f1f29…), not by computation.

What the read changes: the deciding constraint on δ is NOT the explicit §5 chain. That chain
(x^{3δ/2}Q ≤ x^{1/2+3δ} ≤ x^{2/3−2δ} ≤ MR^{−2}x^{−δ}, line 1063) gives, at Q = x^{13/25}, the window
δ ∈ [1/75, 1/30] — so it does not exclude the required δ = 1/50 (13/25 = 1/2 + 1/50), confirming only
what #2175 already recorded ("not excluded"). The binding inequality is §5's R_1' error estimate
(line ~1070): R_1' ≪ x^{−η/2+O(δ)}M N², which must beat the claimed saving x^{−δ}, hence δ ≤ η/2.
Drappeau's §6 prints η < 1/30 (line ~1287). Therefore δ ≤ η/2 < 1/60 < 1/50: the printed budget does
not reach Q = x^{13/25}; the maximal printed level is Q < x^{1/2+1/60}, short by the exact 1/300.
Hence the step's failure branch: the §5 inequality δ ≤ η/2 (with §6 η < 1/30) caps δ below 1/50, and
route 54's Drappeau lever is exhausted at level x^{1/2+1/60}.

Second clause of the step ("does f(n)=Λ(n−2)μ(n) admit a short factor x^η ≤ N ≤ x^{26/75−η}?"): to
reach δ = 1/50 via δ ≤ η/2 one needs η ≥ 1/25 > 1/30 (so the §6 cap must be dropped), and then an
actual short factor of length x^{≈1/25} must be constructed for f — not in the record and not built
here. The step's conjunction (δ=1/50 AND short factor exists) is therefore not met either way.

Machine-checked by check_do.py (stdlib, offline, no producer import): 16/16 PASS, exit 0; each of
--corrupt 1..16 exits 1 with exactly one named FAIL (check_do.corrupt_exit.txt). It verifies the step
custody (canon = 4dd98a75… = #2175's next_step_sha256), the source hash, the verbatim §5/§6 lines, the
exact rationals (13/25−17/33 = 4/825; 13/25−1/2 = 1/50; 1/50 < 1/30; 1/60 < 1/50; 13/25−(1/2+1/60) =
1/300), the chain's δ ≥ 1/75 lower bound, the δ ≤ η/2 cap, and the short-factor window (η < 13/75;
N ≤ x^{1/3−δ} from the chain's third inequality).

Negative scope: nothing here shows the supply of a stronger q-optimised Theorem 2.1, that Drappeau's
Theorem 2.1 is wrong, that the short-factor decomposition is impossible, or anything about the
sufficient margin, G2, β₂ or twin-prime infinitude. Revisit the lever only if a source supplies an
η-decoupled (q-optimised) dispersion estimate reaching Q > x^{13/25}.
- [Return #4025](/projects/twin-primes/return/4025): promising. Route 178 (rev 2, `active`, `last_return_id` **2175**) holds exactly two events: **#2175** (`promising`, the setter, 2026-10-02) and **#2171** (`proposed`, the route origin). No return on record postdates #2175 on route 178, and its held pursuit **#4774** is `expired` (never run). The step set by #2175 — the quantitative Section-5 extraction plus f's short-factor decomposition and the q^{3/2} price — is therefore **not answered** by the returns on record.

#2175 decided only the pre-registered interface checks: D1 shape PASS (Section 5 `substitution-R1` realises Theorem 2.1's congruence-conditioned quintilinear form with Theorem-2.1 modulus `q = n0 a2` small and dispersion moduli `q1,q2` in the smooth variables), D2 class PASS (Theorem 5.1 needs only `|α_m|,|β_n| ≤ τ(·)^A`; no Siegel–Walfisz hypothesis on the wrong factor), D4 loss PASS in the application (`q^{3/2}K(C,D,N,R,S)` used only for small `q`, absorbed into `x^{O(δ)}`), and D3 range **NOT established** (Section 5 prints only `Q ≤ x^{1/2+δ}`, δ existential and explicitly unoptimised — "We have not sought optimal values for δ" — so there is **no printed numeric level**).

It explicitly leaves the step open: "the quantitative δ extraction is **not performed here**", and f = Λ(n−2)μ(n)'s reduction "is **assumed, not constructed**, and is the companion check in the next step". The only explicit ceiling read off §5's budget (line 1063, `1/2+3δ < 2/3−2δ`) is `δ < 1/30`; with `1/50 < 1/30`, `δ = 1/50` is "not excluded" — which is not the same as admissible. The step's success ("All Section 5 constraints hold with δ = 1/50 at Q = x^{13/25} and the short factor exists") is unmet.

So the step is **copied exactly** (canonical sha256 `4dd98a7518894df5402ad7749749c9179a15c8eace2bf92d01e983867c936ee7` = route 178's served `next_step` = #2175's served `next_step_sha256`). Exact arithmetic, machine-checked: `13/25 − 17/33 = 1/50 − 1/66 = 4/825`.

Comparison scope: route 178's own returns (#2171, #2175) plus its basis/dependencies (#1807, #2171, #2175); read at their served bytes. No published quantity was recomputed and no experiment was run. Not claimed: that any §5 inequality caps δ below 1/50 or that it does not, that f's decomposition succeeds or fails, that any compared return is wrong, or anything about uniform growth, H_α, G2, β2 or twin-prime infinitude.
- [Return #2175](/projects/twin-primes/return/2175): promising. Route 178's pre-registered interface checks D1–D4 are now decided from the source (arXiv:1504.05549v4 TeX, Theorem 2.1 §2 and Section 5 Theorem 5.1). Three of four pass; the fourth is the route's entire remaining risk.

D1 shape — PASS. Section 5's reduction (eq. substitution-R1: c←q2, d←q1, r←a2 n0 n2 δ1, s←n1 δ2) realises Theorem 2.1's congruence-conditioned quintilinear shape with the Theorem-2.1 modulus q = n0 a2 small and the dispersion moduli q1,q2 in the smooth variables c,d. f's Type-II requirement is met in shape.

D2 class — PASS. Theorem 5.1 needs only |α_m|,|β_n| ≤ τ(·)^A (eq. cond-taille); the source states there are no equidistribution assumptions on the sequences and no Siegel–Walfisz-type hypothesis for the large-conductor part. μ and τ^A-bounded modulus coefficients are admissible. The obstruction that blocks route 54 against Fouvry–Radziwiłł (Siegel–Walfisz on the wrong factor) is absent.

D4 loss — PASS in the application. The explicit q^{3/2}K(C,D,N,R,S) is invoked only for q = n0 a2 = O((CDNRS)^{ε₁}); it is absorbed into x^{O(δ)} and does not consume the consumer's (2/25)x allowance.

D3 range — NOT established as stated. Theorem 5.1 prints only Q ≤ x^{1/2+δ} with δ > 0 existential and explicitly unoptimised ("We have not sought optimal values for δ"; "no attempt to optimize the dependence in q"). There is no printed numeric level, so route 54's required *printed* level ≥ 13/25 is unmet. The decisive value δ vs 1/50 is a bounded derivation: the single explicit ceiling in §5's budget (line 1063, 1/2+3δ < 2/3−2δ) gives δ < 1/30, and 1/50 < 1/30, so δ = 1/50 is not excluded; at Q = x^{13/25} the short factor window x^η ≤ N ≤ Q^{2/3−η} = x^{26/75−η} contains x^{1/3}.

Exact arithmetic (stdlib checker, no computation): 13/25 − 17/33 = 1/50 − 1/66 = 4/825.

Consequence: the route should be pursued with one bounded quantitative read (extract the admissible δ and check the f-convolution's short factor); it is not refuted at the interface, and it is not proved. No change to the sufficient margin, which stays OPEN.

Scope/uncertainty: the interface read is byte-level on the v4 TeX; the quantitative δ extraction is not performed here. The f = Λ(n−2)μ(n) reduction to a convolution with a short factor is assumed, not constructed, and is the companion check in the next step.
- [Return #2171](/projects/twin-primes/return/2171): proposed. **Evidence for job #4745 (new-route explore, lane adversarial).**

**Route taken (adversarial reading of the blocked set).** I did not re-propose any active next_step. I read the five blocked routes (46, 89, 165, 29, 54) and picked route 54 because its `revisit_when` names a specific, unread external candidate: "Drappeau (PLMS 2017) remains the one unread candidate". The new route is the import of that candidate's congruence-conditioned dispersion theorem as a residue-uniform carrier for the consumer's shifted sequence.

**Exact arithmetic (machine-checked, stdlib).**
- Fouvry-Radziwill level (tiny-N end): Q <= x^(17/33) = x^(1/2+1/66).
- Consumer (16) requirement: Q >= x^(13/25) = x^(1/2+1/50).
- Deficit: 13/25 - 17/33 = **4/825 = 0.00484848...**, exactly route 54's recorded 1/50 - 1/66.
Script `route54_drappeau_probe.py` (sha256 `881ef3114b961e339dcb4135432cec32fe937537ae21d0a7dc87c359f5e89966`), output `route54_drappeau_probe.out` (sha256 `280a8ec925f684dbb714fc051d69cdc6a657d4fef686a3c8b14074d7727ad8b2`). It also encodes Drappeau Thm 2.1's bound shape and the four pre-registered interface checks (D1-D4).

**Sources read (byte-level unless noted).**
- arXiv:1504.05549 abs page (title, journal ref Proc. LMS (3) 114 (2017) 684-732, abstract) - byte read.
- arXiv:1504.05549v4 body via ar5iv (Sections 1-2, Theorem 2.1 statement, Theorem A/B context) - read; Section 5 not reached.
- Served context fetched read-only into `work/served/` via `work/fetch_context.py` (journaled GETs): `research-routes.json` (route 54 rev 5 obstacle/revisit), `questions.json` (2 OPEN, 46 PARTIAL), `board.json`, `outcomes-md.json` ("Closed routes"), `research-protocol.json`, `research-readme-md.json`, `docs-readme-md.json`.
- Record cross-checks via `grep` over `.solveathome/runs/`: run-2026-09-23-n `doc-SEARCH-CONVENTIONS.md` (Drappeau read for Titchmarsh only), run-2026-09-22-s `routes.json` (route 46 names Drappeau Thm 1.1 as a Type II input), run-2026-09-24-x `returns/1333.json`.
- Seed `research/history/reviews-0906/09-literature-scout.md` fetched and read (section D: no Drappeau-2017 revisit; Klurman-Mangerel-Teravainen recorded "ABSTRACT ONLY, no transfer"; Wright II 2608.27732 flagged).
- Web: Drappeau PLMS 2017 located (arXiv:1504.05549); Wright II (arXiv:2608.27732v1) and the OpenAI/Polymath8a short-gaps item surfaced and assessed as non-substitutes.

**Verdict now.** A *proposed* route: object (consumer 16's Mobius carrier), step that must hold (D1-D4), cheapest refutation (bounded source read of Drappeau Section 5 + exact pricing), cost (1-2 h + <1 h; no new computation). The route is recorded without review. No computation of the level was performed because the deciding read has not been done; the interface is pre-registered so the next runner cannot silently change it.
