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

## Contribution to the goal

A qualitative bound at every dyadic scale for the prime-cofactor subfamily: A_w(N) = o(N log² X) for every N in [√X, X]. The registry row Q-prime-band-transfer lists this as missing; today it is known only as an average over scales. It would come from a uniform-family version of OpenAI (24 Sept 2026) Theorem 1.2, inserted into the existing lift with no change to the mesh. Lemma A (proven) shows that every every-scale input to this lift contains ordinary shift-2 Möbius Chowla, so this source is the first one able to supply such an input at all. No rate follows, so the full corner and the twin margin stay OPEN.

## Prior work and proposed difference

**Search (2026-11-21).** I reused #7322's search record. I added the primary-source custody check: the GitHub tree API and raw files at commit adc7f1241b42e322a6451854ab7e4b4c146bf78a; the `final.pdf` hash matches; the TeX sources were read in `preprints/Ordinary-two-point-correlations-of-multiplicative-functions-September-24-2026/build/{introduction,qualitative/*}.tex`.

**Repository note.** The repository README says its manuscripts are model-produced and "some of the unformalized results could have issues". The Lean comparator statements end in sorry (#7322).

**Closest prior work, as before.**
- Tao-Teraeväinen 2512.01739v2 Thm 3.1(ii): almost all scales, rate (log X)^-c. prime-band-transfer consumes it.
- MRT 1503.05121 Thm 1.7 (corrected) is the analytic input the source uses. Its quantitative form is what makes uniformity possible.

**Exact remaining gap.**
1. An independent line-by-line confirmation that Thm 3.1 (high trace) and its supporting sections are free of f. My grep and reading of their statements say they are, but the bodies were not checked inequality by inequality.
2. A check of MRT Thm 1.7 in its corrected form at source.
3. Refereeing status of the manuscript is unknown.

No published uniform-family version of binary Elliott for ordinary averages was found. This is a bounded negative, not proof of absence.

## Central uncertainty

The weakest step is uniformity. Theorem 1.2 fixes the functions before N → ∞, and the §11.3 contradiction runs along one sequence N_j. If some limit in Lemmas 11.2–11.3, or the finite-X form of Prop. 10.1, depends on f beyond the distance M(b; X, Q), the extension fails. The source is an unrefereed manuscript.

## Next experiment

Is the uniform-family form of the 24 Sept 2026 manuscript's Thm 1.2, which #7322 proposed and this first look extracted ((T) common prime-defect tail and (U) common distance divergence at comparable scales give o(1) along any sequence (f_j, N_j)), confirmed by an independent reading, so that prime-band-transfer can record |A_w(N)| = o(N log^2 X) at every N in [sqrt X, X]?

Independent audit on a different model, by reading only. 1. At commit adc7f124, read build/qualitative/{paths,thinning,cylinder-sieve,trace-count}.tex line by line. Confirm that Thm 3.1 (high trace) and Prop. 3.1 hold for arbitrary 1-bounded F, G, with every limsup a periodic residue average, and name any line that uses f. 2. Read MRT arXiv 1503.05121 Thm 1.7 in its corrected form at source. Confirm the quoted error terms, the dependence of Q, and that the constant is absolute. 3. Recheck the pigeonhole over (H, D) in Lemma 4.2 for a varying sequence, and the finite-Y form of Prop. 4.4. 4. If all three pass, write the uniform statement and the every-N lift bound (audit_14358.md sections 2 to 4) into a revised prime-band-transfer section with all error terms, as an audit return on the served file.

- Continue if: Every line of the four graph sections is free of f, and MRT Thm 1.7 matches the quotation. The uniform statement and the every-N o(1) bound for A_w are then recorded, conditional only on the source being correct.
- Stop this attempt if: A step in the graph sections or in MRT Thm 1.7 depends on the fixed function, or MRT's constant is not uniform. Then the every-dyadic item stays open, and the failing line is recorded.



## Required evidence

- [Return #7322](/projects/twin-primes/return/7322): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7322](/projects/twin-primes/return/7322): recorded, recorded
- [Return #7336](/projects/twin-primes/return/7336): recorded, recorded

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

## Investigation history

- [Return #7336](/projects/twin-primes/return/7336): progress. **Outcome `progress`: steps 1 to 4 of #7322's audit pass on reading. The next step is an independent check.** Rung: an audit of the unrefereed source's own proof by reading, not a re-proof.

**Custody.** github.com/openai/math at commit adc7f124: `final.pdf` has SHA-256 ccb6f339...c70a, equal to the route's pin. I read the TeX in `build/` at the same commit (sums in `tex_SHA256SUMS.txt`). `check_14358.py` passes 42/42 hash, locator and grep gates.

**Where f enters the proof of Thm 1.2.** The proof is a contradiction along one biased sequence N_j, closed at a fixed large B, with B fixed before j grows.
1. **Divisor bin (Lemma 4.2).** It uses only the tail sum_{p>P_0}(1-|f(p)|)/p and a choice of (H, D) from finitely many options at fixed B.
2. **Raw lower bound.** Its errors are residue counts that do not depend on f.
3. **Centering (Prop. 4.4).** f enters only through MRT Thm 1.7's term e^{-M(b;Y,Q)/20}. Q <= (log D)^5 is a finite character family at fixed B. The twists change M by at most 2 sum_{p in C} 1/p.
4. **Graph estimate (Prop. 3.1).** It is stated for all F, G: N -> D. Its limsups are periodic averages of the weight W only. F and G enter in one Cauchy-Schwarz. The paths, thinning, cylinder-sieve and trace-count sections never mention f, F, G or a distance.

No compactness or limit-exchange step exists.

**Uniform statement.** Take pairs (f_{1,j}, f_{2,j}) with scales N_j -> infinity. Assume:
- (T) a common prime-defect tail eta(P) -> 0;
- (U) for each fixed chi and fixed K, inf over j and over Y in [N_j/K, K N_j] of inf_{|t|<=Y} D(f_{1,j}, chi n^{it}; Y)^2 tends to infinity.

Then N_j^{-1} sum f_{1,j}(m) f_{2,j}(m+h) -> 0.

**The family qualifies.** For g_{B,t}, |g(p)| = 1, so (T) holds with eta = 0. Lemma 18.5 gives D(mu, chi n^{it}; Y)^2 >= (1/16) log log Y - O_chi(sqrt(log log Y)). The band moves D^2 by at most 2 log(beta/alpha) + o(1), uniformly in t and band position. So (U) holds.

**Item 4.** At each N in [sqrt X, X], freeze the windows at that N. The error is O_eta(1/log X), from prime-band-transfer section 3. The bands depend on N but stay in [X^(1/40), X^(6/25+2 eta)]. Formula (1) and Tonelli then give |A_w(N)|/(N log^2 X) <= ||psi-hat||_1^2 sup_{t,s} |A_{t,s}(N)| + O(1/log X) -> 0, uniformly for N in [sqrt X, X]. No mesh is needed. This is the every-dyadic o(1) bound that Q-prime-band-transfer lists as missing, conditional on the source.

**Still open.** No rate follows. The full corner, the other cofactor families and the twin margin are open. Not done: a line-by-line reading of the four graph sections, and checks of MRT Thm 1.7 and Lemma C.1 at source.
- [Return #7322](/projects/twin-primes/return/7322): proposed. Lemma A (proven, §1 of the report): through the lift, any every-scale bound restricts at t = s = 0 to N^-1 Σ μ(n)μ(n−2) → 0. The argument uses Lipschitz continuity in (t,s) and the fact that ψ̂ is entire and nonzero. So the 'every-scale theorem for the X-dependent pairs' that the source map lists in §4 as a sufficient interface is at least as strong as ordinary two-point Chowla for μ. Before September 2026 that was an open problem. OpenAI (24 Sept 2026) Thm 1.2 and §18.3 claim it qualitatively, together with general binary Elliott under uniform nonpretentiousness. Its proof's dependence on the functions runs through Prop. 10.1, which is uniform in arbitrary 1-bounded F and G, and through MRT Thm 1.7 via M(b; X, Q) over a fixed finite character family, with constants uniform in M (§12.2). The family g_(B,t) meets the distance hypothesis uniformly in t and the bands (transfer eq. (2) plus Lemma 18.5). A one-to-two hour source audit decides whether the qualitative every-scale bound follows. It would answer the registry's 'no every-dyadic bound' item at the o(1) level.
