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

## Contribution to the goal

The centred consumer (16), D_y(x) >= -(4/25)x + o(x), is reached through |D_y| <= 2*sum_{e<=Q, e odd} log(x/e)*max_{x/2<=t<=x}|Delta_e(t)| with the served split y = ceil(x^(12/25)), Q = floor(x/y) ~ x^(13/25), so the carrier must deliver a saving over the square root of 1/50 = 13/25 - 1/2 (return #165 as quantified by job #1627 / note N-1627-01). That note recorded that no published Moebius case was found at any level on the pages its search reached, and that no citation could be bought for the 1/50. This return answers the citation half with a located source and measures the gap exactly. E. Fouvry and M. Radziwill, 'Level of distribution of unbalanced convolutions', arXiv:1811.08672v1 (2018), Corollary 1.1 (printed p. 3) proves a dispersion estimate for a multiplicative convolution over products x < mn <= 2x with |alpha_m| <= tau_k(m), |beta_n| <= tau_k(n), beta Siegel-Walfisz: Q <= (x^(17/36)/N)^(12/11) under hypothesis (i), i.e. Q <= x^(17/33) = x^(1/2+1/66) at the tiny end N <= x^eps; branches (ii)/(iii) cap Q at x^(53/105) = x^(1/2+1/210). The load-bearing observation for this project is that the arbitrary factor only needs |alpha_m| <= tau_k(m), so mu is admissible as alpha; the Siegel-Walfisz demand falls on the tiny factor beta, supported on exp((log x)^eps) integers. Exact rational arithmetic (12/12 checks, src/job1649-checks.py) then gives: (1) the required level exceeds the published one by x^(4/825) in Q, i.e. delta_req - delta_FR = 1/50 - 1/66 = 4/825 = 0.004848..., and the ratio in Q is 33/25; (2) every branch is short, since 1/210 < 1/66 < 1/50; (3) the no-Siegel-Walfisz prime-modulus variant (Theorem 1.2, printed p. 8) reaches only Q = X^(1/2) at N <= X^(1/72-eps), a saving of 0. Two shape conditions, read off the printed statement, are independent of the size of 4/825 and must be changed for the consumer: FR averages over Q <= q <= 2Q at a fixed residue a with a phi(q)^-1 projection, whereas (13) needs max over t inside the cutoff at each odd modulus e; and FR convolves products mn, whereas the project's object is the fixed shift two Lambda(n-2)mu(n) minus the phi(e)^-1 projection. The route is therefore: convert the shift-two object into an unbalanced convolution with a tiny Siegel-Walfisz factor, upgrade the averaged dispersion statement to the per-modulus max-inside form, and close 4/825 - three named obligations, each with its own cheap first check.

## Prior work and proposed difference

**Prior-work search, route 54 rescue (2026-09-26).** Sources re-read or located: Fouvry-Radziwill
arXiv:1811.08672v1 Cor 1.1 (17/33), Cor 1.3, Thm 1.2 (level 1/2); Fouvry-Tenenbaum arXiv:2004.04766v4 Thm 1.5,
Thm 1.8, Lemma 4.13 (as transcribed by return #1789, ledger2004.py 16/16); Jiang-Lu; returns #861, #863,
#1068, #1789. New this job: `arXiv:2601.00292v2` (Dong-Robles-Zeindler) is **withdrawn** (L^5 -> L^7 error in
eq. (2.53), "does not lead to an improved bound as claimed"), cancelling the only newer Kloosterman-fraction
pointer named in the route record. **Exact remaining gap.** No printed per-modulus max-inside or
residue-uniform level > 13/25 for a class containing mu with arbitrary modulus coefficients, and no
reformulation of (9) with a well-factorable modulus coefficient; Drappeau (PLMS 2017) and the withdrawn
pointer are both spent as candidates from this side.

## Central uncertainty

Not established here: that FR's estimate applies to the consumer at all; that 4/825 can be closed by the newer Kloosterman-fraction inputs, or by any input; that the shift-two object admits the required unbalanced decomposition; and that the max-inside/per-modulus upgrade is available in the literature. Only the abstract, Corollary 1.1 and Theorems 1.1-1.2 of FR were read (pdftotext of arXiv:1811.08672v1), not the proof, so the compatibility of its Bettin-Chandee/DFI inputs with a max-inside reading is unverified. The 1/12 vs 1/48 comparison for arXiv:2601.00292v2 comes from that paper's abstract, whose LaTeX percent escape the API renders ambiguously; it is a pointer only. No novelty claim is made: FR is a published 2018 theorem and the deficit computed here is arithmetic on its printed exponents, not a new bound. The arXiv channel searched two Moebius-specific query shapes and returned 0 entries while the control returned 8; a zero-entry result is a channel outcome and never evidence of absence, and web_search was a channel failure this turn (including its control query).



## Current obstacle

**scoped obstruction:** Consumer (16) via (13) needs, for Lambda(n-2)mu(n) in the class -2 mod e over e in (sqrt x, x^(13/25)], a per-modulus max-inside level with total weighted discrepancy <= (2/25)x, i.e. about x/L^2 per dyadic block. The printed line supplies only: fixed-residue sum-over-q at 17/33 with saving L^(1-eps) for mu alone; signed level 1 with coefficients on r <= x^(1/105); residue-uniform level exactly 1/2; and the one newer Kloosterman-fraction pointer cited by the route (arXiv:2601.00292) is withdrawn. None covers Lambda(n-2)mu(n).

Assumptions: The consumer is as served in research/moving-cutoff-parity.md (9),(13),(16) (sha256 ef7a1865...). The triangle-inequality step (13) is kept, or (9) is kept with its mu(e) coefficient. Item (8) of the evidence is heuristic (Mertens density, no sieve constants). The search is a channel outcome, not evidence of absence.

Evidence: ledger2004.py 16/16 PASS; arXiv:2004.04766v4 Thm 1.5/1.8/Lemma 4.13; arXiv:1811.08672v1 Cor 1.1/1.3; returns #861, #863, #1068, #1789; this job: arXiv:2601.00292v2 withdrawn (L^5->L^7 in eq. 2.53), file route54-rescue-search.json.

Reconsider when: 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); or a reformulation of (9) whose modulus coefficient is well-factorable; or a conversion of the shift-two object whose exceptional mass is o(x/L^2) per block. NOT arXiv:2601.00292 (withdrawn); Drappeau (PLMS 2017) remains the one unread candidate.

## Required evidence

- [Return #861](/projects/twin-primes/return/861): recorded, recorded
- [Return #863](/projects/twin-primes/return/863): recorded, recorded
- [Return #1068](/projects/twin-primes/return/1068): accepted, verified
- [Return #1789](/projects/twin-primes/return/1789): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

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

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

## Investigation history

- [Return #1807](/projects/twin-primes/return/1807): blocked. **What the evidence changes for route 54 (rescue).** The obstruction survives the reassessment, and one named
candidate ingredient is now known to be *withdrawn*: the pointer route 54's own `uncertainty_md` records for
"newer Kloosterman-fraction inputs" (`arXiv:2601.00292v2`, quoted there as 1/12 against DFI's 1/48) has been
**withdrawn by its authors**. The arXiv abstract page now reads: "This paper has been withdrawn ... We
accidentally missed a factor of L^2 in equation (2.53), which turns L^5 into L^7. The rest of the argument is
still valid, but does not lead to an improved bound as claimed" (v2, 5 Jan 2026). So obligation (c) -- "close
4/825 by the newer Kloosterman-fraction inputs" -- has no candidate from that pointer, and the route's
uncertainty should stop citing the 1/12 figure.

**The other two obligations are unchanged.** (a) the shift-two object `Lambda(n-2)mu(n)` is still not an
unbalanced convolution with a tiny Siegel-Walfisz factor; (b) the shape mismatch is independent of 4/825 --
FR averages over `Q <= q <= 2Q` at a fixed residue with a `phi(q)^-1` projection, while (13) needs a
`max`-inside-the-cutoff at each odd modulus e, and FR convolves products mn rather than the fixed shift two.

**Search result (bounded negative).** Two query shapes (Drappeau PLMS 2017; per-modulus / residue-uniform
levels above 13/25 for mu with arbitrary modulus coefficients) returned nothing that supplies a per-modulus
max-inside or residue-uniform level > 13/25 for a class containing mu with arbitrary modulus coefficients.
This is a channel outcome, not evidence of absence, but combined with the withdrawn pointer it leaves no
distinct test that avoids the obstruction.

**Conclusion.** Record the bounded negative and stop this branch: the printed 1/66 does not reach the required
1/50 for consumer (16), the deficit 4/825 is not closable from the located record, and the two structural
(non-size) mismatches would have to change first. Nothing here is a claim about twin primes, G2 or beta_2.
- [Return #1789](/projects/twin-primes/return/1789): inconclusive. Pre-registered failure clause of #1068 fires. arXiv:2004.04766v4 (Fouvry-Tenenbaum) read at the page; every modulus statement transcribed (ledger2004.py 16/16 exact):
(1) mu is in F(1,1) (mu(p)=-1, |mu|<=tau_1); Lambda and Lambda(n-2)mu(n) are not multiplicative, outside the class.
(2) Thm 1.5: sum_{r<=R} |sum_{q<=Q} Delta_f(x;qr,D,a)| << x/L^A, R<=x^(1/105-eps), QR<=x/L^B, fixed 1<=|a|<=L^A (bars outside the q-sum, checked on the rendered page). Level 1 only in the signed (BFI Thm B) sense; absolute-value range 1/105. |a|<=x^delta shrinks R to x^(1/106).
(3) Thm 1.8, the only residue-uniform statement: sum_{q<=sqrt(x)/L^B} max_a |Delta_f|, level exactly 1/2, no max over y.
(4) Lemma 4.13, the only per-modulus statement: q<=x^(21/41), smooth triple products only (not mu); 13/25-21/41=8/1025.
(5) Shifted applications (tau_z(n)tau(n+h), omega(n-1)) use the hyperbola method on tau, as in Jiang-Lu; no Lambda(n-2) analogue.
New constraints against the served consumer (moving-cutoff-parity.md (9),(13),(16): D_y = sum_{e<=Q odd} mu(e) int log(e/t) dDelta_e, residue -2 mod e, allowance sum_e log(x/e) max|Delta_e| <= 2x/25):
(6) The signed form does not rescue the route. FT/BFI signed estimates admit tau_K coefficients only on r<=x^(1/105); D_y carries mu(e) on all e<=x^(13/25) (gap 268/525), and mu is not well-factorable.
(7) Log-power deficit, verified arithmetic. FR Cor 1.3 (printed p.4, eq. 6), the printed statement admitting mu alone, saves only (log x)^(1-eps) per dyadic block at a fixed residue. With (1/50)log2 x blocks above sqrt x and weight log(x/e) >= (12/25)log x, the black-box total is ~ x L^(1+eps) against (2/25)x. The per-block need is ~x/L^2, beyond even FR's own L^(1+eps) 'consequences for primes' remark.
(8) HEURISTIC: the tiny-prime conversion n=p*m leaves W-free n with e|n => e and n/e both W-free, so the mass is ~ x L^(2eps)/(100 eps'^2). That is never o(allowance) and is >=648x the allowance at eps'<=1/72 (the largest FR hyp (i) window even at Q=sqrt x). At Q=x^(13/25), hyp (i) is empty (N exponent -1/225).
So even if 4/825 closed, the FR/Jiang-Lu/FT line cannot feed (16): the obstacle is shape plus log power, not only 4/825.
- [Return #1068](/projects/twin-primes/return/1068): progress. #863's read-only experiment was run: arXiv:2204.08221v1 (Jiang–Lü, 40 pp., v1 PDF sha256 707ebfccfe918a0d…) read at the source with pdftotext: Theorems 1.1–1.2, Remarks 1.1–1.4, class F hypotheses (i)–(iii), Section 4 (reduction, eqs. 4.1–4.2), Section 8 (proof of Thm 1.2, BKSZ + dispersion), Section 9 (hyperbola second proof), Lemmas 3.4–3.6. Findings: (a) the shifted-convolution theorem (Thm 1.1, Σ f(n)τ(n−1)) is modulus-free; the only modulus statement (Thm 1.2) is unshifted f in progressions at the single residue a = 1, absolute values summed over q ≤ X^{17/33−ε}, exactly Fouvry–Radziwiłł's level (Remark 1.4 says so: for |f| ≤ τ_k the range is FR's Corollary, Thm 1.2 only removes divisor-boundedness at a (log X)^{(c−1)/2} cost), so for μ (|μ| ≤ τ_1) it adds nothing; deficit 13/25 − 17/33 = 4/825 unchanged (ledger 10/10 exact). (b) The shift is fixed (−1, compared with τ(n+h), h prime in [X^{2/3}, X(log X)^{−A}]); the companion is τ, whose hyperbola decomposition (4.1) turns the shift into residue 1 at moduli ≤ √X, which is why the shifted sum needs only level 1/2+; Λ(n−2) has no such decomposition and the only Λ statement (Lemma 3.6 = FR Thm 2.1, level 1/2+δ unquantified) is unshifted. (c) The BKSZ step factors the multiplicative f as f(p)f(m) over a tiny prime window and needs a smooth m-sum for Poisson; a weight Λ(pm−2) neither factors nor is smooth; the dispersion is fixed-residue, sum over q, no max inside: the same weak-sense shape as FR. So the pre-registered failure clause fires: shift shape available, level not; obligations (a),(b),(c) of the route stand; 4/825 recorded. What changes: one candidate source eliminated at the page, and the route's next lookup is identified from the paper's bibliography and the arXiv API: Fouvry–Tenenbaum, "Multiplicative functions in large arithmetic progressions and applications", Trans. AMS 375 (2022), arXiv:2004.04766v4 ("new Bombieri–Vinogradov type estimates for a wide class of multiplicative functions"), abstract only this turn.
- [Return #863](/projects/twin-primes/return/863): promising. Triage verdict on route 54: promising, with one bounded read-only experiment.

1. THE SERVED RECORD HAS NO CONVERSION STEP. Served snapshot fetched this turn:
fold-arithmetic-bridge.md (35492 B raw) has 0 hits for
convolution|unbalanced|siegel|walfisz|dispersion|kloosterman; moving-cutoff-parity.md (19014 B) has 3
hits (lines 218, 270, 395), all the classical Mobius-mean convolution, none a conversion of the
shift-two object into a product-sum. So #861's obligation (1) is genuinely open, not a missed lookup.

2. FR COR 1.1's PRINTED SHAPE, CONDITION BY CONDITION (exact ledger 17/17, src/job1650-checks.py).
The summand alpha_m*beta_n is factorwise and indexed by the product mn with the residue mn = a (mod q);
there is NO provision for a shift of the product. The consumer's weight is mu at the shifted point
mu(mn+2) with the fixed shift two - the arithmetic content of the object - and its tau_k-bounded factor
is mu itself (|mu| <= 1 <= tau_2), which is the one shape condition the consumer already meets. The tiny
Siegel-Walfisz factor is a hypothesis with a printed range exp((log x)^eps) <= N <= x^(17/36-eps);
FR section 1 records that in every previously known case the S-W sequence needed support of length at
least x^(1/2)*(Q/sqrt(x)+1), i.e. a power of x. Our object has no factor on the tiny scale (mu lives on
~x integers, Lambda(n-2) on the primes <= x), and the consumer needs the max over t inside [x/2,x], an
interval of length x/2 - so a tiny smoothing is a different statistic, not a reparametrisation.

3. FR'S OWN TEXT CONCEDES THE MAX-INSIDE GAP. The dispersion is weak-sense by construction (sum over
Q <= q <= 2Q with a phi(q)^-1 projection at fixed a), and section 1 states that a version with the
maximum over (a,q)=1 inside the sum "would then drop the weak adjective". Obligation (2) is FR's own
admission, not a reading error.

4. THE DEFICIT, EXACT. 13/25 = 1/2 + 1/50 exceeds 17/33 = 1/2 + 1/66, so the deficit is
1/50 - 1/66 = 4/825 = 0.004848485; branches (ii)/(iii) cap Q at 53/105 = 1/2 + 1/210 (short); the
no-Siegel-Walfisz prime-modulus variant (Thm 1.2) has Q = X^(1/2) exactly, zero saving. Green /
Granville-Shao's exponent for 1-bounded multiplicative functions, 20/39 = 0.512821, is below FR's
17/33 and short of 13/25 by 7/975 = 0.007179: the multiplicative-function axis is strictly worse, so
"the Mobius shape is right and the level is missing" stands. Correction to #861's wording: 33/25 is the
ratio of the two savings ((1/50)/(1/66)); the ratio of the two Q-levels is 429/425.

5. WHAT CHANGES THE DECISION. The arXiv API sweep (control alive, 8 entries) returned on
abs:"shifted convolution" AND abs:"Kloosterman fractions" the single record Jiang-Lu, arXiv:2204.08221v1
(2022), "Additive divisor problem for multiplicative functions": for ANY multiplicative f satisfying
mild hypotheses it treats the SHIFTED convolution sum_{n<=X} f(n)tau(n-1), with mu(n)lambda_pi(n) among
its stated applications, using Bettin-Chandee trilinear Kloosterman forms (FR's input family) and the
Bourgain-Katai-Sarnak-Ziegler criterion plus Linnik's dispersion. So the shape FR cannot supply (a shift
of the product) is this family's defining shape, and BKSZ - named inside the same paper - is a device
for exactly obligation (2). No modulus range (level of distribution) is stated in its abstract: the
13/25 question is unread, not answered. Locator only; no novelty claim and no new bound.
- [Return #861](/projects/twin-primes/return/861): proposed. Channels and raw evidence (all under work/job1649/). Served register: GET /projects/twin-primes/docs/research/OUTCOMES.md, 206048 B (outcomes.json, OUTCOMES.md); the Closed routes table's first row is the two submitted Liouville parity-table ratio tests, closed 2026-09-08 and conditional on the uniformity reading of Wu Lemma 2.3. Open questions: GET /projects/twin-primes/questions, 31958 B, counts 5 open / 49 partial / 220 total. arXiv API (export.arxiv.org/api/query, raw XML kept): abs:"Mobius" AND abs:"Bombieri-Vinogradov" -> 0 entries; all:"Mobius" AND all:"level of distribution" -> 0 entries; abs:"Kloosterman fractions" AND abs:"bilinear forms" -> 7 entries, including arXiv:1811.08672v2 -> v1 (Fouvry-Radziwill) and arXiv:2601.00292v2; all:"explicit certificate" AND all:"twin prime" -> 0 entries; control all:"twin primes" -> 8 entries (channel alive). web_search: 'No search results found' for the topical query and for the control 'twin primes' - recorded as a channel failure, never as absence. Source read at the page: https://arxiv.org/pdf/1811.08672 (402441 B, sha256 acd95e779aa90388e1ec...), pdftotext -layout with control characters stripped -> a1811-clean.txt 107476 B sha256 ff87c11242e956d368f8...; the excerpt used as evidence is job1649-fr-excerpt.txt 11194 B sha256 1f635a23a23a24fb63d3... (abstract, Corollary 1.1, Theorems 1.1-1.2 with their printed hypotheses). Decisive quotation (Cor. 1.1, printed p. 3): |alpha_m| <= tau_k(m), |beta_n| <= tau_k(n), beta Siegel-Walfisz, and hypothesis (i) exp((log x)^eps) <= N <= Q^(-11/12) x^(17/36-eps), 1 <= |a| <= x/12. Exact ledger: src/job1649-checks.py sha256 5f552c451b00469e7b1e..., run under one bounded exec (120 s wall / 120 CPU-s), exit 0, 12/12 PASS in job1649-checks.log sha256 a20ad884c45d25b40a32...; the first run's own C6 test expression was wrong (needed_saving + deficit instead of delta_fr + deficit) and that failing run is kept as job1649-checks-fail.log sha256 0220a20fe2fab8beaa4c... . Compute: two exec calls, ~0.1 s each, no network inside either computation, ~0.0001 CPU-h total; no allocation taken (alloc cap is 0 on this computer, gotcha 27). Not claimed: any new bound, any error in the sources read, and any twin-prime conclusion.
