# Review of the completed agent round and its research conclusions **The sufficient twin margin remains OPEN.** Reviewed baseline: `c589d90`, clean working tree, 2026-09-06. The review checks the three most recent round outputs, their mathematical inputs and the conclusions in the live handoff. It is not a reproof of every upstream reduction or every external theorem in the campaign's 79 changed files since `6e4a4fc`. ## 1. Disposition | Finding | Evidence and correction | Consequence | |---|---|---| | R1: the exceptional-scale exclusion was too strong | Section 2 derives a dyadic average from the existing fixed-band continuous estimate, using stability of bounded interval sums and the existing moving-band comparison. | A new theorem valid at every scale is not necessary just to obtain an average over dyadic scales. The saving remains too weak for the twin consumer. | | R2: already controlled proper prime powers were listed as open | corner-correlation §1.1 bounds every term with a proper prime power above V or Z. Section 3 checks the bound including multiplicities. | Keep Q in exact finite diagnostics, but do not reassign its fixed-corner asymptotic bound. The prime-r branches with s>1 and their non-squarefree inputs remain open. | | R3: integer parity was described as the CRT gcd partition | The driver assigns `n % 2`, which computes gcd(n,n-2). Section 4 gives a counterexample to identifying this with a term's modulus gcd or cofactor gcd. | Relabel the output as integer parity. The total C-product and its P/S/Q algebra are retained; no claim of testing the CRT partition follows. | | R4: summary lower bounds exceeded the hypotheses | Zero coefficients or coincident endpoints make the diagonal zero. Report 21 F9 had corrected this in signed-moment but the consolidated summary and reachability note retained the stronger claim. | Retain upper budgets and their limitations; withdraw a universal nonzero diagonal or unavoidable norm floor. | | R5: conditional reachability bookkeeping | The four sets A,B,C,D contain six scalar cutoffs; a fixed-margin bulk need not persist for every gamma<2, though edge strips do. Areas are monotone when gamma is ordered. | Correct six Perron factors and log^6 cost. The conditional gamma=2 threshold for covering the complement of the prescribed corner is retained. No unconditional coverage is added. | The Fourier superposition, the O(1) pretentious-distance comparison, Tonelli order and mesh proof in [prime-band-transfer.md](prime-band-transfer.md) survive the checked hypotheses. I read [Tao–Teräväinen v2, Theorem 3.1(ii)](https://arxiv.org/html/2512.01739v2) and [Matomäki–Radziwiłł–Tao, (1.12)](https://arxiv.org/pdf/1503.05121v3) at source. The technical prime-value condition belongs to case (i). The non-pretentiousness exponent 7/24 permits L=(log X)^(1/4). No deep imported proof is independently reproduced here. The function-class and normalization mismatches in [the source map](next-correlation-source-map.md) remain real. In particular, [Guo v4, Theorem 1.8](https://arxiv.org/html/2608.23500v4) states a logarithmic Liouville-pair bound, not the complex pair needed by the lift. Its proof is not reviewed here. The stronger all-scales input specified in that map is one possible route, not a necessary prerequisite for every averaged or unbounded-scale argument. This review makes no claim about suitable theorems beyond the inspected statements. ## 2. A dyadic sampling consequence missed by the scale discussion All logarithms are natural. Use c from prime-band-transfer (4), decreased if necessary so 0=h/2, and adding gives \[ \frac1{\log X}\sum_{\sqrt X\le2^j\le X}|F(2^j)| \ll \frac1{h\log X}\int_{\sqrt X}^{X}|F(t)|\frac{dt}{t} +B\left(h+\frac1{\log X}+X^{-1/2}\right). \tag{2} \] At most one top interval protrudes beyond X; bound its sampled value by 2B, paying O(B/log X). This retains that endpoint rather than silently discarding it. There are O(log X) sample points. If the integral divided by log X is O((log X)^(-c)), choose h=(log X)^(-c/2). Equation (2) then gives a dyadic average O((log X)^(-c/2)). A measure-zero exceptional set can contain all dyadic points, as the agents observed. It cannot by that observation alone rule out (2): large values of these interval sums persist on neighborhoods whose width is controlled by (1). Arbitrary measurable spikes would lack this property. ### 2.2 Apply the lemma before freezing the moving windows For fixed B,B', apply (2) to \[ a(n)=\mu(n)\mu(n-2)L_B(n)L_{B'}(n-2)/(\log X)^2. \] Set a(n)=0 for n<=2. Its modulus is at most 4 on n<=2X+2. The continuous input is exactly prime-band-transfer (4), uniformly in the bands there. Consequently its dyadic average has saving c/2. Now use that note's exact-window comparison (5). Partition theta=log N/log X into O(1/delta) cells and freeze a band pair on each. The normalized pointwise error between the exact moving-window sum A_w(N) and its frozen-band version is O_eta(delta+1/log X+X^(-1/10)). Its proof counts prime pairs by CRT; the count is a fixed power below every N>=sqrt X. This comparison is pointwise and is therefore applicable at dyadic N. For each cell bound its nonnegative dyadic sum by the sum over all dyadic samples for its own frozen band pair. The O(1/delta) cells cost that factor; no common exceptional set is assumed. We obtain \[ \frac1{\log X}\sum_{\sqrt X\le2^j\le X} \frac{|A_w(2^j)|}{2^j(\log X)^2} \ll_\eta \delta^{-1}(\log X)^{-c/2} +\delta+1/\log X+X^{-1/10}. \tag{3} \] Choose delta=(log X)^(-c/4). Thus, for some fixed d=c/4=c_*/2>0, \[ \boxed{\frac1{\log X}\sum_{\sqrt X\le2^j\le X} \frac{|A_w(2^j)|}{2^j(\log X)^2} \ll_\eta (\log X)^{-d}.} \tag{4} \] The campaign's upper endpoints are x=2N, so dyadic N also gives dyadic x. The proper-square error in converting A_w to the raw P-by-P contribution is O(N^(19/20+epsilon)); divided by N log^2 X and averaged it is negligible for a sufficiently small fixed epsilon. No eta depending on X is used. **Payoff and limit.** (4) is a dyadic scale-average relative saving, and in particular supplies an unbounded sequence of dyadic scales with that relative saving for this subfamily. It is not a bound at every prescribed dyadic point, o(N), or a positive margin. The resulting averaged bound on |A_w|/N is O_eta(log^(2-d) X); d is small and unextracted. Other prime-r cofactor branches and the complement of S_0 remain unestimated at the required precision. No change to W_dagger or E_dagger follows. ## 3. Proper prime powers already have a fixed-corner bound For W>=2, \[ \sum_{p^a>W,\ a\ge2}p^{-a}\ll W^{-1/2}. \tag{5} \] Split p at sqrt W. For each smaller prime, the tail starting at the first eligible exponent is at most 2/W; for larger primes it is at most 2/p^2. Bounding primes by integers proves (5). There are at most x times this sum integers n<=x divisible by one such power. For each n the number of divisor/cofactor choices and all logarithmic weights are O_epsilon(x^epsilon), after distributing a sufficiently small fixed epsilon among the divisor bounds. Thus the total absolute contribution of all Q-containing products is \[ O_\epsilon(x^{22/25+\epsilon}+x^{39/40+\epsilon}), \tag{6} \] as already derived in corner-correlation §1.1. Taking epsilon<1/40 absorbs this into every fixed O_H(x/log^H x). Repeated powers count with their multiplicity; the divisor bound pays for it. This does not remove non-squarefree inputs arising from small repeated prime factors in the remaining P/S products, and does not bound the complement. ## 4. What the finite driver measures The exact integer prime-log-vector comparison tests the full C coefficient, including all P/S/Q branches. It survives this review and reruns. Its `n % 2` split is exactly gcd(n,n-2), hence integer parity. It does not compute the per-term gcd of the two CRT moduli, or that of the two cofactors K,T in corner-correlation (1). For example, 330=5*11*6 and 328=41*2*4 give nonzero Mobius divisors d=5,e=41, prime powers r=11,r'=2 and s=6,s'=4. Then gcd(330,328)=2 but gcd(dr,er')=gcd(55,82)=1; likewise the pair d=10,e=41,K=33,T=8 has gcd(K,T)=1 while n is even. Both occur in the toy proxy window x=512, U=V=Dlo=4, Y=Z=Elo=1, D0=102, E0=255; 330 lies in J_x. These are proxy lower cuts, not a claim that the asymptotic S_0 is populated at x=512. A future CRT-branch test must retain the relevant indices before grouping. The total identity keeps all these terms; the existing output is relabeled instead of claiming a partition that it never computed. The retained corner probe is also empty because Elo=48854902 exceeds E0=44739242 at the recorded parameters. This blocks its right side over the entire dyadic interval, independently of the left-prefix restriction. The proxy matrix remains useful for finite algebra only. ## 5. Further corrections to the live summaries The general diagonal is nonnegative but only bounded above by a constant times B²C²f²MN/A under the stated hypotheses. b=0, c=0, an empty harmonic set or equal endpoints refute a uniform positive lower bound. The sqrt(M) norm cost is an upper bound, not an unavoidable lower bound for every actual coefficient. These were already identified in review 21 F9. The consolidated unsigned-mass comparison also mixed a single dyadic box with the full fixed-eta prime-cofactor subfamily: the stated log^(K+2) excess belongs to the latter's x log^2 x envelope, as corner-correlation §1.5 derives. That envelope is not a lower bound on the signed sum. In reachability (1), let t=2-gamma+2eta_0. In the bulk p,q>=2eta_0, both p+3q and 3p+q are at least 8eta_0. Therefore that bulk's open uncontrolled part is empty when gamma>=2-6eta_0. Edge strips still persist for gamma<2: p=2eta_0,q=0 fails the right cross bound and the left zero bound. The claim that the bulk and both strips all persist for every gamma<2 was too strong; the stated overall threshold stays 2. The four conditional controlled sets in reachability §4.4 contain six scalar inequalities. At most six Perron integrals, with O(log^6 x) cost, are required for their intersection. The uniform twists already available cover them; the claimed four-factor separation omitted the vertical cutoffs. Area is monotone for nested sets: 37/25<3/2 explains why the table's 37/25 row is larger despite being printed later. Ineffective constants mean no effective numerical onset has been obtained, not that an asymptotic onset does not exist. The consolidated claim is corrected accordingly. No necessity or exhaustion conclusion is inferred from these bookkeeping corrections. ## 6. Verification and next decision [round-review-validation.js](round-review-validation.js) checks finite interval stability, exact integration of the fixed-sequence sampling bound, endpoint handling and counterexamples to dropping the neighborhood factor, using arbitrary complex test sequences rather than prime data. It also checks the parity/gcd counterexamples, prime-power-tail bounds, the fixed-margin bulk counterexample and all six cutoff exponents. The driver tails and existing transfer validators are rerun through the custody tools. These tests validate finite algebra and catch specified errors; they do not prove a logarithmic correlation rate. Verification run: `node research/qc.js --index --strict` (zero findings), `node research/qc/selftest.js` (all known positives and controls pass), `node research/audit-numbers.js` (all checks pass), and `git diff --check`. The new sampling validator, transfer-review validator, readiness validator and corner-correlation validator reproduce their embedded outputs. Both branch-driver tails were regenerated after label corrections; their measured coefficient figures reproduce. QUESTIONS and SCRIPTS are regenerated. These gates do not independently prove the imported source theorems or the mathematical conclusions of this review. Retain the agents' coefficient identities, Fourier lift and source matching work with these corrections. The next arithmetic priority is a matched estimate for the remaining prime-r terms with s>1 or s'>1 and the full one-sided/global payoff, or a substantially stronger averaged rate. Requiring a theorem that excludes every dyadic exceptional point would unnecessarily discard the route in §2. Another local kernel rectangle alone still would not complete the global argument.