# Average the whole identity before isolating the corner **Twin-prime infinitude remains OPEN.** This is a change of representation and a derived one-point cancellation statement, not a signed improvement. It revises the priority of extending the isolated corner. No claim is made that it is the best possible proof route or a novel identity in the literature. Baseline: clean commit 3a15541, 2026-09-06. The cross-check used the current question/outcome registers, the arithmetic chain, the earlier [cross-campaign review](cross-campaign-synthesis.md), the proposal registry, and the owning arguments listed in section 6. It is not a fresh independent audit of every archived manuscript, script or imported theorem. ## 1. The distinction that changes the next question Smoothing only the isolated corner leaves a sharp-to-smooth transition and E_out. Both are open at the required signed precision. The sharp transition has a non-negligible squared norm for sufficiently small fixed eta. Instead, the initial Vaughan cutoffs are choices in an identity for the **same** twin count. Average those identities while their already-proved Type I estimates hold uniformly. The resulting full smooth remainder represents the original count; there is no need to restore a chosen sharp corner. The arithmetic previously in E_out is included, not discarded. This is a different operation from averaging the corner cutoffs D_L approximately x^(19/25-2eta) and D_R approximately x^(19/20-2eta). The cutoffs averaged below are the much smaller initial U and Y. ## 2. A uniform family of complete reductions Keep x dyadic, J_x=(x/2,x], V=floor(x^(6/25)), Z=floor(x^(1/20)), and C2 as in [prime-detection-spec.md](prime-detection-spec.md). Choose integers \[ a_L=\lfloor x^{11/50}\rfloor,\quad b_L=\lfloor x^{6/25}\rfloor, \qquad a_R=\lfloor x^{1/25}\rfloor,\quad b_R=\lfloor x^{1/20}\rfloor . \tag{1} \] For real u in [a_L,b_L] and v in [a_R,b_R], use the exact cuts d>u, e>v in the two Vaughan expansions, leaving V,Z fixed. For integer divisors these are the cuts floor(u), floor(v). Define \[ C_{u,V}(m)=\sum_{\substack{d\mid m\\d>u}}\mu(d)\beta_V(m/d), \qquad R_{u,v}(x)=\sum_{n\in J_x}C_{u,V}(n)C_{v,Z}(n-2). \tag{2} \] Then, for every fixed A>0, \[ S(x)=C_2x+R_{u,v}(x)+O_A(x/\log^A x), \tag{3} \] **uniformly over the whole rectangle of u,v.** Here is the uniformity check against the owning proofs, rather than an assumption that any choice of cutoff works. | part of the existing reduction | largest support or smallest length | fixed margin | |---|---|---| | first prime BV Type I sums | floor(u)V <= b_L V <= x^(12/25) | below x^(1/2) by exponent 1/50 | | comparison term with b(n) | nonempty Mobius intervals have length scale T=x/k > a_L | the uniform mean in polylog-fold-transfer section 4 applies at endpoints at least a_L, a fixed power of x | | second Mobius BV Type I sums | introduced modulus <= b_R Z <= x^(1/10); T=x/k > a_L | sqrt(T) is at least a constant times x^(11/100), leaving exponent margin 1/100 | | nonprimitive congruences | the same gcd 1, gcd 2 and odd-lift cases as before | the bounded parity reduction and intervals T or T/2 leave those fixed margins intact | The coefficient bounds used in the first and second Type I sums remain |c(d)|<=log d. Their endpoint maxima already include the varying lower cuts; summing beta_V(k) x/k costs only fixed logarithms. Thus the arbitrary logarithmic precision of each imported estimate absorbs the same losses uniformly, with constants depending on A and the fixed exponents in (1). No x-dependent epsilon or shrinking power margin is used. The proofs being extended are [prime-detection-spec.md](prime-detection-spec.md) sections 2–4, [polylog-fold-transfer.md](polylog-fold-transfer.md) section 4 and [shifted-prime-decomposition.md](shifted-prime-decomposition.md) sections 2–4. The latter's Mobius BV input has its own owning derivation in [mobius-bv-derivation.md](mobius-bv-derivation.md); it is not assumed from a citation to a prime BV theorem. The primary Vaughan identity and prime BV statement were rechecked in [Tao, Notes 3, Lemma 18 and Theorem 17](https://terrytao.wordpress.com/2015/01/10/254a-notes-3-the-large-sieve-and-the-bombieri-vinogradov-theorem/). Their complete analytic proofs are not newly audited here. ## 3. The averaged global coefficient and exact bookkeeping For i=L,R put L_i=log(b_i/a_i) and \[ \rho_i(d)= \begin{cases} 1,&d\le a_i,\\ \log(b_i/d)/L_i,&a_iu) du/u. Independent probability measures on the two cutoff intervals therefore give \[ \langle R_{u,v}(x)\rangle_{u,v} =\sum_{n\in J_x}G_L(n)G_R(n-2)=:\mathcal R(x). \tag{5} \] Averaging (3), using its uniform error, yields \[ \boxed{S(x)=C_2x+\mathcal R(x)+O_A(x/\log^A x).} \tag{6} \] The main term is unchanged because the left side of each averaged identity is the same S(x). This is not a positive average over translations of the primes, and it does not assume that good sets of x intersect. All divisors, nonsquarefree integers, prime powers inside beta, strict cuts and shifted endpoints are retained. There is no separate T or E_out in (6). Relative to the existing fixed-cut reduction, mathcal R=E_dagger+O_A(x/log^A x); this is a comparison of the complete signed sums, not a termwise deletion of E_out or an extension of the previous controlled geometric region. For another exact check, if rho is any cutoff profile, write mu_rho(d) =mu(d)rho(d) and mu_h=mu-mu_rho. Dirichlet convolution gives \[ \Lambda=\Lambda_{\le W}+\mu_\rho*\log -\mu_\rho*\Lambda_{\le W}*1 +\mu_h*\Lambda_{>W}*1. \tag{7} \] Indeed log=Lambda*1 and mu*1=delta, so the right side collapses to Lambda_(<=W)+Lambda_(>W). The analytic uniformity above requires the specified profiles and supports; (7) alone proves no estimate. ## 4. What the full coefficient norm buys Put F_i(m)=sum_(d|m) mu(d)rho_i(d). Exactly, \[ G_i(n)=-\sum_{\substack{r\mid n\\r>W_i,\ r=a_i and L_i>=1, sum_(a_ib_i, so the corresponding part of (4) has h_i(d)=1. Thus \[ G_i=C_i+B_i,\qquad B_i(m)=\sum_{\substack{d\mid m\\d\le D_i}} \mu(d)h_i(d)\beta_{W_i}(m/d). \tag{10} \] On I_L=J_x and I_R=J_x-2 the owning sharp estimate has ||C_i||_2^2=Theta_eta(x log^2 x), while (9) has ||G_i||_2^2=O(x log x). Hence, by Cauchy, \[ \langle C_i,B_i\rangle =-\|C_i\|_2^2+\langle C_i,G_i\rangle =-\|C_i\|_2^2+O_\eta(x\log^{3/2}x). \tag{11} \] In particular the ratio of this cross term to ||C_i||_2^2 tends to -1. Much of the corner's one-point energy is canceled by the rest of the same full coefficient. Treating the corner as the whole problem can discard that cancellation. **This is not a shift-2 assertion.** Neither (11) nor its norm proof estimates C_L(n)B_R(n-2), B_L(n)C_R(n-2), or mathcal R. Replacing a common input by two shifted inputs is precisely the unproved step. ## 5. The next attempt and what would count as success The subsequent [smooth-majorant note](global-smooth-majorant.md) gives an absolute O(x) bound for a C3 probability profile inside the same uniform rectangle. This is a different admissible pair; only the complete signed residual is transferred to arbitrary logarithmic precision. The signed or sufficient-constant estimate remains OPEN. Research judgment: prioritize a bounded signed attempt on the global pair G_L(n),G_R(n-2), allowing the two initial profiles to vary within explicitly admissible supports. Keep cofactor sums inside the coefficient until a justified inequality is applied. The new small-cofactor transfer remains a possible ingredient; it is not automatically applicable to G_i, which is neither 1-bounded nor multiplicative. The consumer is a one-sided lower bound: for some fixed c,K>0 on an unbounded common set of dyadic scales, \[ \mathcal R(x)\ge -C_2x+\frac{c x}{\log^K x}. \tag{12} \] Equation (6), with A>K, then implies twins. A justified signed scale average can also serve; retain its rate and positivity implication. The stronger mathcal R=o(x) would give the full weighted main term and is not a necessary research requirement. The [first factor-sign attempt](global-factor-signs.md) now gives an exact small-prime formula, pays its exceptional prime-power contribution, and refutes a pair-trigger negative-part majorant. Continue from its full signed cofactor sum and explicit prime-filter correction to seek an estimate at scale x. The positive part may have to be retained: an unsigned negative-part bound is sufficient, not necessary. A norm bound of O(x log x), or a relative logarithmic saving too small to reach scale x, does not meet (12). No choice of profile is certified to do so. The immediate falsifiers are a failed uniform BV margin, a changed main term when averaging, an omitted r=n correction, or an unretained piece in (10). These have analytic derivations above and finite controls below. An eventual signed attempt must separately test its proposed inequality on admissible factor configurations and price all its terms. ## 6. Why this precedes the other routes in the current corpus | candidate | checked reason for its current priority | |---|---| | Extend small cofactors term by term | The two priced losses in cofactor-progression-transfer remain: power-sized moduli and accumulated cofactor cost; its restricted rate is still weaker than o(x). | | Shrink eta with x | reachability-coverage section 2.5 leaves x^epsilon constants unpaid; shrinking geometry does not control its signed mass. No uniform repair is supplied here. | | Smooth only the isolated corner | The norm transition is non-negligible; sections 2–4 above instead average the whole identity in an admissible initial-cutoff range. | | Use a better prime-detecting weight | chen-opportunity-audit identifies a real removable penalty, already removed in prime-detection-spec. The present use of profile freedom stays within that complete consumer; a generic optimization is not itself a new estimate. | | Return to anchored survival or fold transport | The anchored survivor implication is valid, but no new anchor-specific lower bound is supplied. The exact transport evaluator does not supply the missing recurrence. Those missing inputs are not refuted just because this arithmetic route is pursued first. | | Demand generic Type I/II control | Ford–Maynard's source theorem concerns specified uniform coefficient classes. The actual-coefficient estimate here does not imply those hypotheses; their prime-free examples are not a universal obstruction to this task. | Sources checked for the last two comparisons: the relevant proofs and limits in [paper/anchored-note.md](../paper/anchored-note.md), [the transport proposal](../paper/proposals/prop-tailcount-transport.md), [chen-opportunity-audit.md](chen-opportunity-audit.md), and [Ford–Maynard v1, Theorem 2.1](https://arxiv.org/html/2407.14368v1). No proposal grade, novelty assessment or publication status is changed. [global-cutoff-averaging-validation.js](global-cutoff-averaging-validation.js) checks exact finite coefficient identities using integer prime-log vectors, independent cutoff averaging, the complete two-expansion product identity, and controls for omitted prime powers, r=n and lower-divisor complements. No finite output is a test of the asymptotic estimates or a twin margin. The finite fixture uses rational averages of integer cutoffs, an exact finite counterpart of (5) and (7), rather than pretending to test the asymptotic exponent range at a small x. The analytic extension and the same-input comparison have been self-audited here; independent mathematical review remains appropriate before relying on them downstream. Verification on 2026-09-06: the finite validator passed; strict QC with index regeneration reported zero findings; the verifier's 58 positive and 47 negative controls passed; the numerical audit passed all 251 checks; git diff --check was clean. The embedding tool initially required an output banner; a later descriptive label correction was re-embedded with its recorded override and no numerical change. These mechanical checks do not certify the analytic derivation.