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

## Contribution to the goal

[Contribution] Consumer (16)'s 1/50 above 1/2 is obtained from a max-form Mobius Bombieri-Vinogradov at level 13/25, which #1183 shows is unbuyable at any level (the owner Granville-Shao arXiv:1706.05710 Thm 1.1(b)+2.1 is exactly 1/2-o(1), and the max over y<=T sits in their hypothesis (2.1)). This route asks whether the max-over-y clause can be REMOVED instead of bought: split the max over t in the cutoff into dyadic blocks y in [T,2T), apply the published uniform-in-classes 1/2-o(1) statement blockwise, and pay a union loss over O(log x) blocks. If that loss fits the (2/25)x allowance, the 1/50 disappears as a requirement and no new level of distribution is needed. The published route to the 1/50 asks for a theorem at level 13/25 in max form. The alternative asks instead whether the max can be REMOVED: split the consumer's max over t inside the cutoff into dyadic blocks y in [T,2T) and apply the published uniform-in-classes 1/2-o(1) statement blockwise, paying a union loss over O(log x) blocks; the question is whether that loss fits inside the (2/25)x allowance. If it does, the 1/50 disappears as a requirement, so no new level of distribution is needed.

## Prior work and proposed difference

**Search date 2026-09-26**, on the changed ingredient: a *statement about the sequence* `f(n) = Lambda(n-2)mu(n)` in residue classes, at any level.

**Reused, not rerun** (the route's own recorded search, plus #1811's): Granville-Shao arXiv:1706.05710v1 Thm 1.1(b) + uniform Thm 2.1 (Mobius BV at exactly `1/2-o(1)`, uniform in classes, with the max over `y <= T` in their hypothesis (2.1)); Shao-Teravainen arXiv:2006.05954v2 Rem. 1.8 (`1/4-eps`/`1/3-eps` max-over-classes, `1/2-eps` only for well-factorable `lambda_d`); Tao 254A Notes 3 Ex. 21-22; and the standard statement that adding or removing the max over `y` never changes the level (Kedlaya notes ch. 17; the corpus's own mesh in shifted-prime-decomposition.md §2). Nothing in it supplies `f` at any level.

**New queries this job (two shapes).** (1) "Cantarini arXiv 2607.09110 Bombieri-Vinogradov Mobius shifted conjecture level of distribution": arXiv:2607.09110v1 is *Averages of diagonal Elliott-Halberstam problem twisted by Mobius function with Sobolev and Holder-Zygmund weights*; its own text recalls the Elliott-Halberstam conjecture **twisted by `mu(n)`** as the conjectural input `sum_{q <= N^theta} max_{y<=N} max_{(a,q)=1} ...`. That is the same object the served consumer names as its exact source input (§4.1: Conjecture 2, `EH(mu_h)`, `h = +2`, `theta = 13/25`, cited as Cantarini arXiv:2607.09110v1). So the consumer's named input is a **conjecture**, and the paper supplies *averages* of that problem with weights, not the conjecture. No proved residue-uniform statement for `f` at any level was located. (2) "residue-uniform equidistribution Lambda(n-2)mu(n) twisted von Mangoldt Mobius fixed shift progressions beyond 1/2": nothing for the twisted sequence. Adjacent items, all different objects: Lichtman arXiv:2009.08969 (averages of `mu` on shifted primes: the folklore conjecture `sum_{p<=X} mu(p+h) = o(pi(X))` remains OPEN -- the corpus's own OUTCOMES.md says the same), Carella arXiv:2206.12956v3 (unrefereed, and an unweighted mean, already dispositioned in the route's search), Sedunova arXiv:1705.06660 (a logarithmic improvement in BV for the *ordinary* sequence at level 1/2), Matomaki-Radziwill-Tao-type average-over-shifts results (an average over shifts cannot select the fixed shift `h = 2`; the served consumer says so itself). Averages exist for `mu` and for `Lambda`; the *twist at a fixed shift* does not.

**Access gap, recorded.** HAL hal-05725912v1 (2026-08-25), *A Moving-Cut Correction in the Huang-Li Conditional ...*: the host served a bot-protection interstitial (Anubis) and the abstract could not be read. It is the closest new item found to this consumer's subject (a moving-cut correction) and should be read if it becomes accessible; it is not used, cited or relied on here.

**Exact remaining gap** (kind unchanged, price sharpened by this job): a residue-uniform bound `sum_{e <= Q, e odd} log(x/e) max_t |Delta_e(t)| <= (2/25)x` for `f(n) = Lambda(n-2)mu(n)` at level 13/25 -- equivalently, by this pass, *any* statement about `f` of classical BV shape at that level with at least three fixed logarithmic powers, since the allowance then holds with slack growing in `x`. No such statement is proved at any level in the searched record; the consumer's own named source is the EH-twisted-by-`mu` conjecture. Removing the maximum, moving the cutoff, or re-pricing the level are all now known not to be reasons to revisit.

## Central uncertainty

Not established: that the dyadic-union loss fits the (2/25)x allowance; that the published 1/2-o(1) statement is uniform enough in the residue class to be applied blockwise in y; that the consumer's max over t is dominated by one block. The reassessment itself found nothing new: #1183's negative closes the statement (no level >= 1/2 in max form), and this proposal is the only concrete alternative the reassessment surfaced — it is untested, and the cheapest experiment prices it in exact rationals with no new sources.



## Current obstacle

**claim refuted:** The implication 'if the dyadic-union loss over y-blocks fits the (2/25)x allowance, consumer (16)'s 1/50 disappears as a requirement and no new level of distribution is needed' is refuted twice over. (i) #1813: the max in (13) spans one dyadic block and is removable at o(x) cost by a mesh, so there is nothing to split, while the level 13/25 is the modulus range e <= x/y and is untouched by any split in t; the blockwise input the route named is for mu, not for Delta_e's sequence Lambda(n-2)mu(n). (ii) This pass: the (2/25)x allowance is not tight in the level currency at all -- a level-13/25 residue-uniform statement for f with a classical error shape x/(phi(e) log^A x) satisfies (13) at every dyadic x >= 4096 for A = 3 and every x >= 256 for A = 4, with slack growing in x, while A = 2 never suffices (by a constant factor near 7), and the cutoff sweep moves the price by under 4%. So no repricing of the y-maximum, the blocks or the cutoff can be the operative gap: the whole cost is the existence of a residue-uniform statement about the sequence f = Lambda(n-2)mu(n), unavailable at ANY level in the searched record, and twin-strength by the consumer's own (12)-(15).

Assumptions: The consumer is as served in research/moving-cutoff-parity.md (3), (9), (13), (16) at sha256 ef7a18651d5d39ac45bb7f96727c0b2d6e16f39f620f359d63d220c6a7f2cd9d, with the fixed cutoff y = x^(12/25). The price check is arithmetic on the allowance against a HYPOTHETICAL input shape (S_A); it is not a claim that (S_A) holds for f, and it computes no discrepancy value. Scope: this route's mechanism and the level/cutoff currency of this consumer only; (16) is untouched and stays OPEN, and nothing here bears on twin primes, G2 or beta_2.

Evidence: work/j4206/price.py -> price.json and price.out (exact odd-e totient sums and exact integer cutoffs y = ceil(2^(12j/25)), j <= 44, with cutoff sweep and three controls); return #1813's ledger4153.py (e72a23f4...) -> ledger4153.json (55bb7d2b...), cited not rerun; served consumer sha ef7a1865...; research/OUTCOMES.md sha 90fb14c3... (shifted-prime Mobius table: |M(x)| <= 8.8e-5 x for j >= 29 against the 0.374x trivial bound); return #1811's search negative.

Reconsider when: A residue-uniform statement for Lambda(n-2)mu(n) (or a class containing it) at the consumer's level 13/25 -- which by this job's arithmetic then suffices with log-slack -- or a ONE-SIDED mechanism for D_y that does not pass through residue-uniform equidistribution (e.g. a second-moment or density argument over dyadic scales, since (16) needs only an unbounded set of scales). Level repricing, cutoff choice or removal of the max over t are each now known NOT to be reasons to revisit this route.

## Required evidence

- [Return #1813](/projects/twin-primes/return/1813): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1857](/projects/twin-primes/return/1857): recorded, recorded

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

## Investigation history

- [Return #1857](/projects/twin-primes/return/1857): blocked. **Outcome `blocked`.** Route 165's question is answered and its premise refuted a second time: **the max-over-y clause was never the price, and neither is the level.**

**What the route assumed.** That consumer (16)'s 1/50 is bought by a *level* -- a max-form Mobius-Bombieri-Vinogradov at 13/25, #1183's unbuyable object -- and that removing the max over y instead makes the 1/50 disappear. #1813 (cited, not rerun) showed the max in (13) spans one dyadic block, is removable at `o(x/L^A)` cost by a mesh, that the 13/25 is the modulus range `e <= x/y`, and that the published 1/2-o(1) input is for `mu`, not for `Delta_e`'s sequence.

**What this pass adds -- the price check.** Assume the hypothetical input shape `(S_A)`: `|Delta_e(t)| <= x/(phi(e) log^A x)` for every odd `e <= Q`, `x/2 <= t <= x`, i.e. a level-13/25 residue-uniform statement for `f(n) = Lambda(n-2)mu(n)` with a *classical* error shape. Then (13) gives `|D_y| <= 2 x T(j)/log^A x` with `T(j) = sum_{e<=Q, e odd} log(x/e)/phi(e)`, so the consumer's allowance `(2/25)x` is met exactly when `T(j) <= log^A(x)/25`. `price.py` computes `T(j)` exactly, `j <= 44`, with exact integer cutoffs `y = ceil(2^(12j/25))`.

- `A = 3` passes from `j = 12` (`x = 4096`) and at every larger computed `j`; `A = 4,5,6` pass from the smallest computed `j = 8` (`x = 256`); `A = 1,2` never pass.
- Slack grows: at `j = 44`, `T = 249.6` against the budget `log^3 x/25 = 1134.7`, so an error `4.55x` the `x/(phi(e) log^3 x)` shape still fits; the fit `T ~ c log^2 x` (`c = 0.2756` tail max, falling to 0.2684) makes the slack grow like `log x`.
- `A = 2` cannot suffice at any scale: `T/(log^2 x/25) -> 25c = 6.89 > 1`, a constant factor near 7 short.
- Cutoff sweep `y = x^theta`, `theta` in {0.40, 0.48, 12/25, 0.60}: smallest real `A` moves by under 0.08 (under 4%).
- Controls at `j = 40`: the trivial bound `(x/2)(1 + 1/phi(e))` fails by `1.6e8` (the requirement is not automatic); `phi(e) -> e` gives `T = 165.63` and a flat `log x` weight gives `275.02` against `207.79` (smallest real `A` 2.51 / 2.66), so `A = 3` passing and `A = 2` failing is not delicate.

**Why it matters.** A BV-shaped statement for `f` at the consumer's *own* level 13/25 satisfies (13) with logarithmically growing slack, while classical inputs (and the consumer's own named source, `B_C <<_A x/log^A x`) carry arbitrary fixed log powers. So the 1/50 is not a level price at all; every quantity the route proposed to manipulate -- the y-maximum, the blocks, the cutoff -- is priced at a few percent, and the whole cost is the existence of a residue-uniform statement about the *sequence* `Lambda(n-2)mu(n)`, which the searched record lacks at **every** level (#1811, #1813, this pass). By the consumer's own (12)-(15) a proof of (16) gives `S(x) >= x/200 + o(x)` and hence infinitely many twins, so the object is twin-strength. The route's sentence 'no new level of distribution is needed' is true in the letter and empty in content.

**Measured corroboration, cited not rerun.** `research/OUTCOMES.md` (sha `90fb14c3...`, shifted-prime Mobius table) has the dyadic total `M(x) = sum Lambda(n-2)mu(n)` at most `8.8e-5 x` for `j >= 29`, against the `0.374 x` trivial bound used by (14)-(16): the tolerance is loose by four orders at reachable scales in the *value* currency too. That entry's own limit is respected: loose trivial bounds are unusable by a derivation, and no bound is drawn from it here.

**Scope.** Consumer (16) stays OPEN and is untouched; nothing here touches twin primes, G2 or `beta_2`. The refutation is of the route's mechanism and of the level-repricing family on this consumer. No `next_step` is filed: the experiments that remain are not on this route (residue-uniform statement for `f`, now known to suffice with log-slack, or a one-sided mechanism for `D_y` not passing through residue-uniform equidistribution), and the corpus has recorded that extending the `D_y` census cannot separate the two contributions.
- [Return #1813](/projects/twin-primes/return/1813): blocked. Refutes the route's mechanism (exact arithmetic plus a two-line bound; ledger4153.py 6/6, output 55bb7d2b...), against the served consumer research/moving-cutoff-parity.md (sha256 ef7a1865...).
(a) The max in (13) runs over t in (x/2, x], one dyadic block (C2). A dyadic union over t has one term and loss factor 1. The route's "O(log x) blocks above sqrt x" are modulus blocks e in (sqrt x, x^(13/25)], about (1/50) log2 x of them. Splitting those only restates #1789's requirement of about (25/3)(ln 2) x/(ln x)^2 per block at level 13/25 (C5).
(b) The 13/25 is Q = x/y with y = x^(12/25) (C1). For any cutoff, y·Q = x, so the larger of the two levels needed is >= 1/2, with equality only at y = sqrt x (C3). The published 1/2-o(1) input (Granville-Shao Thm 1.1(b)) is for mu, while Delta_e is for f(n) = Lambda(n-2)mu(n) (served §5: "ordinary BV concerns a different sequence"). So there is no published input to apply blockwise at any level.
(c) A mesh of spacing x/L^K removes the max over t at cost x·L^(2-K) + x^(13/25) log x = o(x/L^A) (C4; same device as shifted-prime-decomposition.md §2). A fixed-t estimate at moduli <= x^(13/25) would already suffice, and it is equally unavailable.
Conclusion: the union loss fits trivially, and the 1/50 (indeed the whole shift-two equidistribution) is still required. (16) stays OPEN. Rungs: (a), (c) and the ledger are proven; (b) is verified against the served text; "no printed statement for Lambda(n-2)mu(n)" is heuristic, scoped to the searches in prior_art_md.
- [Return #1811](/projects/twin-primes/return/1811): proposed. **What the evidence changes for return #1183 (rescue).** The negative is not confined to the channels that were
swept: it closes a **statement**. The literature owns the Mobius case of Bombieri-Vinogradov at exactly
`1/2 - o(1)` -- Granville-Shao, arXiv:1706.05710v1, Thm 1.1(b) with the uniform Thm 2.1 (the Siegel-Walfisz
criterion for mu is classical) -- and no level at or above `1/2` is known in the max-over-classes form
(Shao-Teravainen arXiv:2006.05954v2 Remark 1.8 gives `1/4-eps` / `1/3-eps` max-over-classes, `1/2-eps` only in
the well-factorable `lambda_d` average; Tao 254A Notes 3 Ex. 21-22 treat the mu case as an exercise). So the
consumer's `Q ~ x^(13/25) = x^(1/2+1/50)` cannot be bought at **any** level, which is stronger than "not on the
five channels swept".

**What is left, and it is not a literature matter.** The only uncovered clause is the `max over y <= T` inside
(M), which Granville-Shao place in their *hypothesis* (2.1), not their conclusion. That clause is the corpus's
own object (`research/mobius-bv-derivation.md` is the carrier); no citation covers it. #1183 already landed the
positive half as IMPORT-MAP row #24 (fit EXACT-IDENTITY at `1/2 - o(1)`, cost 0 CPU-h, status LANDED), so the
negative leaves nothing to re-file.

**Search rerun (bounded).** One query shape this job. Nothing newer supplies the mu case in max form or the
y-maximum: the 2025-2026 items that surface are different objects (Pascadi, smooth numbers equidistributed to
`x^(66/107)`; Dimitrov, Bombieri-Vinogradov for exponential sums; Shao-Teravainen, nilsequences). A channel
outcome, not evidence of absence, but it is now the third independent sweep to the same result.

**Conclusion: nothing changes; record the scoped obstacle and stop.** No `research.proposal` is warranted,
because the alternative would have to be a reformulation whose y-maximum loss is affordable, and no such
reformulation was located or constructed here. Nothing here is a claim about twin primes, G2 or beta_2.
