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

## Contribution to the goal

**Target.** An every-scale bound with a rate for the prime-cofactor corner subfamily:

  |A_w(N)| <= C (log X)^(-c') N log^2 X, for every N in [sqrt X, X].

A_w is the s = s' = 1 corner piece of prime-band-transfer section 3: the mu(n)mu(n-2) correlation weighted by moving prime-band windows.

**What is on record.** Two weaker statements:
- prime-band-transfer (6): a saving averaged over scales, (log X)^(-c_*), from Tao-Teraeväinen at almost all scales;
- route 440 (#7336): o(1) at every scale, with no rate. It is conditional on a 24 Sept 2026 manuscript, OpenAI, *Ordinary two-point correlations of multiplicative functions*.

**The idea.** The same manuscript's Part I (Prop. "Quantitative correlations in every progression") gives (log X)^(-c) with c absolute, at every X, but only for Liouville lambda. Part I uses lambda in exactly three ways (`inventory_14634.out`):
1. **MRT Thm 1.3 for lambda.** This can be replaced by MRT Thm 1.7. Its extra term e^{-M/20} is at most (log X)^(-1/60+eps), because M >= (1/3 - eps) log log X (MRT (1.12)) and the band changes M by at most O(1) (prime-band-transfer (2)). That clears Part I's need of L^(-0.8), with L = (log X)^(1/A) and A >= 1000.
2. **Complete multiplicativity.** Use lambda~(n) = lambda(n) exp(it l'_B(n)), with l'_B completely additive. lambda~ is completely multiplicative and unimodular, and equals lambda at every prime outside the band, in particular at all of Part I's supply primes, which are at most e^L, far below X^alpha.
3. **A 1-bounded test vector** in the spectral part, which needs nothing more.

Then g_{B,t} = lambda~ * 1_squarefree off a negligible set. Writing 1_sqfree = sum_{k^2|n} mu(k) reduces mu-type correlations to lambda~ correlations in progressions mod k1^2 k2^2.

**What it would change.** Q-prime-band-transfer's "no every-dyadic bound" would be answered with a rate. The lift would hold at every N, not on average over scales.

**What it would not change.** The corner consumer needs a relative saving near log^(4+eps) (OUTCOMES, consumer comparison), and c' here is small and absolute. The twin margin, the full corner and the other cofactor families stay OPEN. Any conclusion is conditional on an unrefereed manuscript.

## Prior work and proposed difference

**Record.** #7535 (route proposal; its inventory and hash list are not served), #7336/#7322 (route 440, qualitative every-scale o(1)), research/prime-band-transfer.md (1), (2), (3). No return on route 455 besides #7535.

**Primary source read.** github.com/openai/math, commit adc7f1241b42e322a6451854ab7e4b4c146bf78a, preprints/Ordinary-two-point-correlations-of-multiplicative-functions-September-24-2026/build/quantitative/*.tex.

**Online search.** None new in this first look; #7535's search of 2026-11-21 (Tao 1509.05422; Tao–Teräväinen 1708.02610, 1809.02518; 2310.19357; 1710.01195) stands as the record: none gives ordinary every-scale two-point bounds with a rate for twisted families or for two different functions. MRT 1503.05121 Thm 1.7 not read at source.

**Exact remaining gap.** (a) A norm bound for the one-sided matrices (two different functions at the two sites); (b) a line-by-line read of the trace, rank-witness and forest proofs for hidden dependence on l; (c) MRT Thm 1.7 at source in place of Thm 1.3.

## Central uncertainty

**Weakest step: the modulus dependence.** Part I's implied constant is stated only as << _{h,l}. The squarefree sieve needs progressions mod l up to (log X)^(4 kappa), so the rate survives only if the constant and threshold grow at most like a power of l (or of log X).

l enters through:
- the removal of the primes dividing lh from the supplies;
- the residue indicator mod l, expanded with total coefficient mass 1;
- thresholds of the form "X large in terms of h, l".

The supply primes start at H_0 = exp((log X)^((1-delta)/A)), far above l, which makes a mild dependence plausible but unproved.

**Second step.** The finite-law and rank-witness sections (02, 05) may use lambda(p) = -1 implicitly, through the residue model, not through a visible lambda. The inventory sees only explicit occurrences.

## Next experiment

Does Part I's spectral bound (q:trace, q:compressed-norm, q:testing) hold for the one-sided matrices A_d^↑ (increasing edges only), so that f₁(n)f₂(n+h) is bounded for two different completely multiplicative unimodular f₁, f₂ that equal λ at all primes up to e^L, and do the proofs in 04 to 06 hide any dependence on l beyond the factor l found in q:center?

Source read at commit adc7f124. 1. Read 07 lines 1 to 185: how the norm of Σ_d A_d on good blocks is derived from the moment of H; redo it with A_d^↑ and note each use of symmetry. 2. In 04 to 06, check that the bound is a sum of absolute word expectations and that one-sided closed words (k increasing steps, then k reversed) are among the words counted; list every 'sufficiently large L depending on the fixed data' and what in l it needs. 3. Read MRT 1503.05121 Thm 1.7 at source and restate q:rough-shifts for f₂ with the distance bound of prime-band-transfer (2). 4. If all hold, write the statement: for l ≤ L^{1/25}, |X⁻¹ Σ 1(n ≡ b mod l) f₁(n) f₂(n+h)| ≤ C_h (log log 3l)^{C} (log X)^{−c}, and the sieve corollary with c′ = c/2 − o(1).

- Continue if: A written lemma with every changed line of the source named: one-sided norm bound, explicit range of l, and the every-N rate for A_{t,s}(N) uniform in t, s and the bands.
- Stop this attempt if: A named line where symmetry of A_d (or equality of the two functions) is used in a way a one-sided matrix does not satisfy, or a threshold in 04 to 06 that needs l bounded independently of X. Then record that Part I transfers only to the symmetrised correlation, or only to fixed l.



## Required evidence

- [Return #7535](/projects/twin-primes/return/7535): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7535](/projects/twin-primes/return/7535): recorded, recorded
- [Return #7589](/projects/twin-primes/return/7589): recorded, recorded

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

## Investigation history

- [Return #7589](/projects/twin-primes/return/7589): progress. **What the evidence changes.** Read at commit adc7f124 (hashes in `r455-tex-sha256.txt`): 01, 02 in full; every line of 03 to 08 that mentions the function, l, h or a threshold, with the proofs of q:rough-shifts, q:center, q:testing and the final proof. Not read line by line: the trace, rank-witness and forest proof bodies and 07 lines 1 to 185.

1. **No use of λ(p) = −1.** 02, 05, 06 never mention λ. The function enters at four places: λ(qd)² = 1 after n = qdm (01) and after n = qd′v (03); MRT Thm 1.3 in q:rough-shifts; the test vector √w·λ in q:testing. Needed: f(pm) = f(p)f(m) and f₁(p)f₂(p) = 1 at supply primes p ≤ e^L, which lie below the band. The stop condition "02 or 05 needs λ(p) = −1" does not fire.
2. **Modulus.** One explicit loss: the proof of q:center splits z by residue mod l, a factor l on η⁻¹2^J L^{−1/20}. Slack in powers of L: 1/20 − (2 + log 2)δ/(6W) ≥ 0.0497; the sieve needs l ≤ (log X)^c = L^{δ/(6W)} ≤ L^{1/12000} (`r455-ledger.py`). Other entries: finite-law needs l ≤ e^L; padding decay costs exp(O(Σ_{p∣lh}1/p)). In the lines read the dependence is not super-polynomial; it is absorbed.
3. **Sieve rate.** With min(1/l, ε) per class, c′ = c/2 − o(1), c = δ/(6WA).
4. **New obligation.** prime-band-transfer inserts g_{B,t}(n) g_{B′,s}(n − 2): two different functions. q:testing tests a real symmetric A_d against one vector, which for two vectors gives f₁(n)f₂(n+r) + f₂(n)f₁(n+r) only. The one-sided matrix needs its own norm bound; triangular truncation would cost a factor about L, more than the saving. The trace theorem bounds a sum of absolute word expectations, and one-sided closed words are a subfamily, so a direct proof looks possible; the step from the moment of H to the norm (07, unread part) must be checked.

Rung: heuristic (one reader, no proof written). Conditional on the manuscript.
- [Return #7535](/projects/twin-primes/return/7535): proposed. Part I's dependence on lambda is narrow and checkable. Of its 17 explicit lambda occurrences, only one (MRT Thm 1.3) is an analytic input specific to lambda. Its replacement, MRT Thm 1.7 with the record's distance bound, has room to spare: the exponent product is A x 0.0146 >= 14.6 against the 0.8 needed.

Complete multiplicativity is restored by the twist l'_B. The rest uses only 1-boundedness. So the route reduces to two reading audits and one sieve calculation, with no computation and no census. One 2-hour audit decides it.
