# A two-class Erdős–Rankin lower bound from finite estimates **Local exposition for manuscript-binding review, 8 October 2026.** Chris Benjaminsen provided project direction. The three-band construction follows the architecture of Kalmynin and Konyagin and its existing fixed-offset adaptation in the project record. Earlier mathematical sources and contributors retain their credit. This new exposition was prepared by the actual gpt-6.1-sol assistant at high reasoning effort from the frozen Lean sources. It makes no claim to sole authorship or novelty. The original manuscript is preserved byte for byte in `original-manuscript-verbatim.md`. The preserved contribution record distinguishes historical authorship from this revision. The original project manuscript's Methods and AI disclosure records [return #1093](https://solveathome.org/projects/twin-primes/return/1093) under nielsegberts with model gpt-6-astra (the manuscript names GitHub Copilot CLI), and [return #1730](https://solveathome.org/projects/twin-primes/return/1730) under Benjaminsen with deepseek-v4-flash (the manuscript names Freebuff). It also retains [return #158](https://solveathome.org/projects/twin-primes/return/158) for the historical OCR/source-custody and priced-constant discussion, and the recorded Codex revisions. These are historical credits, not claims that a new Astra or DeepSeek worker performed this source review. The linked [project contribution record](https://solveathome.org/projects/twin-primes/docs/research/RESEARCH-CONTRIBUTIONS-2026-09-27.md) and complete original disclosure remain part of the provenance; no author dates have been reconstructed here. ## Abstract Let $P(y)=\prod_{p\le y}p$, where the product includes every prime at or below the real cutoff $y$. The integers surviving the two exclusions are $$ \mathcal T_y=\{n\in\mathbb Z:\gcd(n(n+2),P(y))=1\}. $$ We explain the frozen formal proof that their largest consecutive gap satisfies $$ G_2(P(y))\gg \frac{y(\log y)^3(\log\log\log y)^2}{(\log\log y)^4} \qquad(y\longrightarrow\infty). \tag{1} $$ The proof constructs a cover of a finite interval by pairs of residue classes, bounds its remaining indices, and assigns a different reserved prime to each remaining index. The estimates needed for this construction are proved at the fixed parameter $A=12$. A finite Selberg sieve, a shifted finite Euler product, and an elementary factorial argument supply the three counting budgets. In particular, the final proof has no prime-number-theorem or smooth-number-asymptotic hypothesis. We also explain why the finite natural-number definition of $G_2$ is exactly the attained maximum over all integer starts, including negative starts and gaps crossing a period boundary. Seven stronger or more general claims printed in the original manuscript remain separate research obligations. Their original statements and domains are retained below. The mathematical exposition here is a new manuscript binding; evidence for the unchanged Lean objects does not by itself review this new text. ## 1. The theorem and the integer gap it measures All logarithms are natural. For a finite set $S$ of primes, put $P_S=\prod_{p\in S}p$. An integer $s$ starts a consecutive survivor gap of distance $d\in\mathbb N$ when $$ d>0,\qquad s,s+d\in\mathcal T_S, \qquad s+i\notin\mathcal T_S\quad(00$ and $Y$ such that, for every real $y\ge Y$, an attained largest integer survivor gap $d\in\mathbb N$ exists for $S=\{p\in\mathbb N:p\text{ prime},\ p\le y\}$, where the last comparison uses the real value of $p$, and $$ c\frac{y(\log y)^3(\log\log\log y)^2}{(\log\log y)^4}\le d. $$ “Largest” means both that some integer start realizes distance $d$ and that every consecutive survivor gap at every integer start has distance at most $d$. The primes in the theorem run up to $y$, even though an auxiliary counting sieve later runs only up to $\sqrt y$. The formal statements are `PrimaryMain.manuscript_main_goal`, `IntegerMainBinding.integer_maximum_binding`, and `IntegerMainBinding.manuscript_integer_main_goal`. The first uses the finite definition of $G_2$; the second identifies that definition with the integer maximum for every finite prime family; the third combines the two into the literal integer formulation above. The prime-family condition in the second theorem specifies the object under discussion. It supplies no analytic estimate. The formal parameters are explicitly `(S : Finset Nat)` and `(hS : PrimeFamily S)`, where `PrimeFamily S` means that every member of S is prime. Thus `assumptions: []` in the claim inventory means no additional analytic inputs or custom axioms; it does not erase this defining parameter/domain condition. Intermediate positivity and range hypotheses are discharged by the parameter lemmas before the final eventual claims are used. Here is the identification in detail. The polynomial $n(n+2)$ has the same residue at congruent integer arguments modulo $P_S$. Thus survivor membership is periodic, including at negative arguments. The residue $P_S-1$ survives: both $P_S-1$ and $P_S+1$ are coprime to $P_S$. This also covers the empty family, where $P_S=1$ and every integer survives. Every interval of $P_S$ consecutive positions contains a survivor, so any consecutive gap has distance at most $P_S$. The finite definition considers starts $0\le sX$. Thus the actual weights satisfy $$ \sum_{d\mid t}w_t=\frac{Xg(d)}d+r_d, \qquad |r_d|\le g(d). \tag{8} $$ The sifted weight sum is exactly $S(X,\Omega)$. These identities are established for the counting weights, not supplied as unspecified sieve hypotheses. Put $\nu(p)=g(p)/p$, extending multiplicatively on squarefree divisors, and $h(d)=\prod_{p\mid d}\nu(p)/(1-\nu(p))$. At a real level $H>1$, the Selberg denominator is $$ J(H)=\sum_{\substack{d\mid P_{\mathrm{aux}}\\d^2\le H}}h(d). $$ The explicit coefficients are $$ \lambda_d=\nu(d)^{-1}J(H)^{-1}\mu(d) \sum_{\substack{k\mid P_{\mathrm{aux}}\\d\mid k,\ k^2\le H}}h(k). $$ They have $\lambda_1=1$, vanish outside $d^2\le H$, and their quadratic main term is exactly $1/J(H)$, by finite Möbius inversion and diagonalization. Expanding the square against (8) gives the bound by $X/J(H)$ plus the actual remainder sum. The construction uses the squared-level convention above throughout. The remainder is controlled coarsely but explicitly. One has $|\lambda_d|\le d$ on supported divisors. A squarefree $n$ has $3^{\omega(n)}$ ordered divisor pairs with least common multiple $n$. Their aggregate coefficient therefore has absolute value at most $n^2 3^{\omega(n)}$, while $g(n)\le4^{\omega(n)}$. Only $n\le H$ contributes. The factors at $2$ and $3$ give $$ 12^{\omega(n)}\le4n^2, $$ so the entire remainder is at most $4H^5$. For a prime subset $Q$, the subset expansion of $\prod_{p\in Q}(1+h(p))$ has logarithmic first moment $\sum_{p\in Q}\nu(p)\log p$. If this is at most $\log H/4$, at least half the mass has $d^2\le H$. Consequently $J(H)\ge\tfrac12\prod_{p\in Q}(1-\nu(p))^{-1}$. A finite prime-log estimate provides $$ \sum_{p\le N}\frac{\log p}{p}\le\log4\,(2+\log N),\qquad N\ge2. \tag{9} $$ It follows from the elementary primorial bound $\theta(x)\le x\log4$ and finite summation, and does not use the prime number theorem. Choose $k=512\log4$, $N=\lfloor\exp(L/k)\rfloor$, and $H=\exp(L/16)$. These satisfy the required moment condition on primes up to $N$. Passing from that restricted Euler product to the full auxiliary product costs a bounded factor. The proof bounds its logarithm by $$ E=20\log4\,k+80\log4. $$ Also $4H^5\le y/(12L)$ on the stated domain. With $C_s=2e^E$, the resulting finite estimate is $$ S(X,\Omega)\le C_sX V(\sqrt y)+\frac y{12L}, \qquad V(\sqrt y)=\prod_{p\le\sqrt y}(1-g(p)/p). \tag{10} $$ The product is bounded by elementary finite estimates as well. Define $C=4\exp(10\log4)$. The ordinary inverse product obeys $$ \prod_{p\le N}(1-1/p)^{-1}\le C\log N\quad(N\ge2), $$ and the harmonic sum gives the complementary upper bound for the ordinary product. Splitting a prime band therefore bounds its product by $C\log N_0/\log N_1$. At primes at least $3$, the basic two-class factor is at most the square of the ordinary factor; middle-band primes supply two additional powers. The exceptional prime $2$ costs at most a factor $2$. Converting real cutoffs to floors, with the proved logarithmic floor bounds, gives $$ V(\sqrt y)\le128C^2 \frac{(\log z_0/\log z_1)^2}{L^2}. \tag{11} $$ Set $K=C_s\,128C^2$. Substitution of (3), using the natural floor bound, yields $$ S(m,\Omega)\le\left(\frac{K12^4}{B}+\frac1{12}\right)\frac yL. $$ The actual choice is $$ B=1+6K12^4, $$ which is positive, at least $1$, and gives $S(m,\Omega)\le y/(4L)$. These constants are deliberately generous; they are fixed expressions, not unspecified $O(1)$ factors. The proof establishes a sufficient one-sided estimate. It does not assert the original two-sided product asymptotic. ## 6. The smooth-number budget at $A=12$ For any $\alpha>0$, $X\ge1$, and $Z\ge2$, the finite Rankin inequality is $$ \Psi(X,Z)\le X^\alpha \prod_{p\le Z}(1-p^{-\alpha})^{-1}. \tag{12} $$ Indeed, each smooth $1\le n\le X$ has $1\le X^\alpha n^{-\alpha}$. Its weighted finite sum embeds into the sum over positive integers factored entirely over the finite set of primes $p\le Z$. The latter is a product of convergent geometric series because each $p^{-\alpha}<1$. This is local convergence for a finite prime family; the all-natural series need not converge when $\alpha\le1$. The identification with the actual smooth count uses `Nat.smoothNumbersUpTo X (⌊Z⌋₊ + 1)`, preserving the inclusive condition $p\le Z$ despite that library's strict bound convention. Let $U=\log X/\log Z\ge1$, suppose $\log Z\ge1$ and $U\le\sqrt Z$, and set $$ \delta=\frac{\log U}{2\log Z},\qquad\alpha=1-\delta. $$ Then $0\le\delta\le1/4$, $\alpha>0$, and $p^\delta\le\sqrt U$ for every prime $p\le Z$. The shifted local ratio satisfies $$ \frac{1-1/p}{1-p^{\delta-1}} =1+\frac{p^\delta-1}{p-p^\delta} \le\exp\!\left(\frac{4(p^\delta-1)}p\right). \tag{13} $$ For this bound, $p^\delta\le\sqrt p\le3p/4$ makes the denominator positive and at least $p/4$; then $1+t\le e^t$ applies. The inequality $e^x-1\le xe^x$ gives $$ p^\delta-1\le\delta\log p\sqrt U. $$ Summing (13)'s exponent with (9), and using $2+\log Z\le3\log Z$, bounds the total shift cost by $6\log4\sqrt U\log U$. Combining the product with the ordinary inverse bound and $$ X^{1-\delta}=X\exp(-U\log U/2) $$ proves $$ \Psi(X,Z)\le CX\log Z \exp\!\left(-\tfrac12U\log U+6\log4\sqrt U\log U\right). \tag{14} $$ Now use the actual $X=m+2,Z=z_1,U=u$. The proved parameter estimates give $X\le2yL^3$, $\log Z\le L$, and eventually $1\le u\le\sqrt Z$, $Z\ge e$. They also give the quantified limits $$ \frac{\log X}{L}\longrightarrow1,\qquad \frac{\log u}{L_3}\longrightarrow1,\qquad \frac{u\log u}{L_2}\longrightarrow12, $$ and $u\longrightarrow\infty$. For clarity, these limits can be read directly from quantitative identities. Put $r=\log X/L$. The floor and mass estimates give $1\le r\le1+7/\sqrt L\le2$, and $$ u=r\frac{12L_2}{L_3},\qquad \log u=L_3+\log12-\log L_3+\log r. $$ The bound $\log t\le2\sqrt t$ makes both $r-1$ and $(\log u)/L_3-1$ arbitrarily small at explicit sufficiently large thresholds. Their product gives $(u\log u)/L_2\to12$. Moreover $u\ge\sqrt{L_2}$ proves its growth. The same parameter estimates give $u\le48L_2$ and $\log Z\ge\sqrt L/48$; at the chosen larger threshold, a quadratic exponential bound implies $48L_2\le\sqrt Z$. Thus the range in (14) is proved for these exact parameters. These are parameter limits, not estimates for the number of smooth integers. Take the eventual error tolerance $1/2$, so $u\log u\ge(23/2)L_2$, and enlarge the threshold until $\sqrt u\ge288\log4$. Then the positive shift term in (14) is at most $u\log u/48$. Its total exponent is therefore at most $$ -\frac{23}{48}u\log u\le-\frac{529}{96}L_2\le-\frac{11}{2}L_2. $$ The factor $X\log Z\le2yL^4$ and $\exp(-11L_2/2)=L^{-11/2}$ give $$ \Psi(m+2,z_1)\le\frac{2Cy}{L\sqrt L}. $$ Finally enlarge the threshold until $\sqrt L\ge96C$. It follows that $$ 2\Psi(m+2,z_1)\le\frac y{24L}. \tag{15} $$ Every range and growth condition used here is discharged in `SmoothBudget.eventual_smooth_and_uncovered_budget`. Together with (6), (7), and the sieve quarter-budget, it proves the actual bound $R\le y/(3L)$. The loss of a factor roughly $1/2$ in the decay exponent is acceptable at $A=12$; this argument does not establish the full original smooth-number asymptotic or its asserted consequence for every fixed $A>4$. ## 7. Enough primes in the reserved interval We next prove the needed literal supply $$ \#\{p\text{ prime}:y/2\le p\le y\}\ge\frac y{3\log y}. \tag{16} $$ This is derived by a factorial argument. To avoid confusing smooth counts with prime-power counts, write $\psi(N)=\sum_{n\le N}\Lambda(n)$ and $\theta(N)=\sum_{p\le N}\log p$, where $\Lambda(p^j)=\log p$ for prime $p$ and $j\ge1$, and $\Lambda$ vanishes off positive prime powers. Put $f(x)=x\log x-x$, with $f(0)=0$. Logarithmic ratio inequalities show, for positive integers $n$, $$ \log n\le f(n+1)-f(n)\le\log(n+1). $$ Induction gives, for $N\ge1$, $$ N\log N-N+1\le\log(N!) \le N\log N-N+1+\log N. $$ For $F(x)=\log(\lfloor x\rfloor!)$, this implies the uniform estimate $$ |F(x)-f(x)|\le1+\log(x+1)\quad(x\ge0). \tag{17} $$ For $x\ge1$, let $n=\lfloor x\rfloor$. Logarithmic ratio inequalities give $(x-n)\log n\le f(x)-f(n)\le(x-n)\log x$. Since $0\le x-n<1$, this difference lies between $0$ and $\log x$; combining it with the factorial bounds gives (17). The range $0\le x<1$ is handled separately, where the factorial is $0!=1$ and $-1\le f(x)\le0$. Thus applying (17) at $x/30$ does not require $x\ge30$. Define the signed floor wheel and the factorial combination $$ w(n)=n-\lfloor n/2\rfloor-\lfloor n/3\rfloor -\lfloor n/5\rfloor+\lfloor n/30\rfloor, $$ $$ T(x)=F(x)-F(x/2)-F(x/3)-F(x/5)+F(x/30), \qquad a=\frac7{15}\log2+\frac3{10}\log3+\frac16\log5. $$ Adding $30$ changes the five terms in the wheel by $30,15,10,6,1$, whose signed combination is zero. Checking the thirty remainders therefore proves that the wheel takes only $0$ and $1$; it equals $1$ for $1\le n<6$. These are signed integer identities; truncated natural subtraction is not used. The exact combination of the five $f$-terms is $ax$, so (17) gives $$ |T(x)-ax|\le5(1+\log(x+1)). \tag{18} $$ The divisor identity $\sum_{d\mid n}\Lambda(d)=\log n$, followed by finite summation, gives $$ \log(N!)=\sum_{n\le N}\Lambda(n)\lfloor N/n\rfloor. $$ Substituting it in the factorial combination yields the exact identity $$ T(N)=\sum_{n\le N}\Lambda(n)w(\lfloor N/n\rfloor). $$ Nonnegativity and the wheel properties now imply $$ 0\le T(N)\le\psi(N),\qquad \psi(N)-\psi(\lfloor N/6\rfloor)\le T(N). \tag{19} $$ With $L_N=1+\log(N+1)$, strong induction proves $$ aN-5L_N\le\psi(N) \le\frac65aN+5L_N^2. \tag{20} $$ For the upper bound at $N\ge6$, take $M=\lfloor N/6\rfloor$. The induction hypothesis and (19) bound $\psi(N)$ by $aN+5L_N+(6/5)aM+5L_M^2$. Since $M\le N/6$, the linear terms are at most $(6/5)aN$. Also $N+1\ge3(M+1)$ and $\log3\ge1$, so $L_M+1\le L_N$. Hence $L_N^2-L_M^2\ge L_N$, absorbing the new error. For $N<6$, the elementary bound $\psi(N)\le N\log(N+1)$ handles the base case. The higher-prime-power tail satisfies, for $N\ge2$, $$ 0\le\psi(N)-\theta(N)\le2\sqrt N(\log N)^2. \tag{21} $$ Indeed, every nonprime prime power $n\le N$ is $p^j$ with $p\le\lfloor\sqrt N\rfloor$ and $2\le j\le\lfloor\log N/\log2\rfloor$. This finite set lies in the image of those base-exponent pairs. Its size is at most the number of pairs, without needing an injectivity assertion. Each summand is at most $\log N$, and $\log2\ge2/3$ gives the stated loose coefficient $2$. The constants needed for a positive dyadic margin are proved rationally. The inequalities $(24/23)^{16}\le2$, $(12/11)^{12}\le3$, and $(8/7)^{12}\le5$, together with $\log r\ge1-r^{-1}$, give $$ \log2\ge2/3,\qquad\log3\ge1,\qquad\log5\ge3/2, \qquad a\ge31/36. $$ Also $a\le2$ follows from $\log t\le t-1$ at $2,3,5$. For real $y\ge4$, use $N=\lfloor y\rfloor$, $M=\lfloor y/2\rfloor$. Combining (20) and (21), with $N\ge y-1$ and $M\le y/2$, proves $$ \theta(y)-\theta(y/2)\ge\frac{2a}{5}y-\mathcal E(y), $$ $$ \mathcal E(y)=a+5L_++5L_+^2+2\sqrt y(\log y)^2, \qquad L_+=1+\log(y+1). $$ There is a concrete error threshold. If $y\ge300000^4$ and $\log y\ge1$, put $s=\sqrt{\sqrt y}\ge300000$. Then $y=s^4$, $\sqrt y=s^2$, and $\log y=4\log s\le8\sqrt s$. Therefore $(\log y)^2\le64s$, $L_+\le10s$, and $L_+^2\le576s$. It follows that $$ \mathcal E(y)\le128s^3+2932s\le3060s^3\le y/90. $$ The last inequality uses $90\cdot3060=275400\le s$. Since $2a/5\ge31/90$, the difference of theta sums is at least $y/3$. This difference sums $\log p$ over the strict interval $y/20$, with the fixed $B$ of Section 5. On the positive large domain the floor mass in (3) equals $c\cdot yL^3L_3^2/L_2^4$. The elementary floor inequality gives $$ c\frac{yL^3L_3^2}{L_2^4} **Input M.** Mertens' prime harmonic estimate is > $$ > \sum_{p\leq t}\frac1p=\log\log t+M+o(1). > \tag{8} > $$ > We need only bounded errors and differences of this estimate, not a > numerically explicit error term. Its use is recorded in [R4, section 6.3] > and in [KK, section 2]. ### O2. OPEN — Uniform smooth-number asymptotic (original Eq. 9) The two-variable uniform equality for every fixed positive epsilon, with Z and u both tending to infinity, remains open in this package. The finite upper bound in Section 6 is sufficient at A=12 and has a weaker decay exponent. Original statement and domain (`original-manuscript-verbatim.md`, lines 171–184): > **Input H ([HT, Corollary 1.3, printed page 417]).** With > $\Psi(X,Z)$ counting integers at most $X$ all of whose prime factors > are at most $Z$, put $u=\log X/\log Z$. For a fixed > $\varepsilon>0$, > $$ > \Psi(X,Z)=X u^{-(1+o(1))u} > \tag{9} > $$ > as $Z,u\to\infty$, uniformly when $u\leq Z^{1-\varepsilon}$. > The formula and its range were read from the primary page image: this is > the printed range of Corollary 1.3 (p. 417). In $x,y$ the source restates > the underlying range (1.13) of its Theorem 1.2 as > $(\log x)^{1+\varepsilon}\leq y\leq x$, (1.13)$'$ on p. 418. > We will use only its upper-bound consequence. ### O3. OPEN — The arbitrary fixed A>4 route (original Eqs. 23–24) Quantified parameter limits underlying Eq.23 are proved for $A\ge4$ with $B=\operatorname{chooseB}(A)$. The printed smooth equality and $o(y/L)$ conclusion for every fixed $A>4$ in Eq.24 are not proved. The main theorem fixes $A=12$; its logarithmic shape is unchanged. Original statement and domain (`original-manuscript-verbatim.md`, lines 389–413): > ### 6.3 The smooth-number term > > Set $X=m+2$, $Z=z_1$. Then > $$ > \log X=L+O(L_2),\qquad > u=\frac{\log X}{\log Z}\sim\frac{A L_2}{L_3}. > \tag{23} > $$ > Both $u$ and $Z$ tend to infinity. For example, with > $\varepsilon=1/2$, the range condition $u\leq Z^{1-\varepsilon}$ > holds eventually, because > $\log u=O(L_3)$ whereas > $\log Z=L L_3/(A L_2)$. > > Moreover $u\log u=(A+o(1))L_2$. Input H therefore implies > $$ > \Psi(m+2,z_1) > =(m+2)L^{-A+o(1)} > =o(y/L)\qquad(A>4). > \tag{24} > $$ > For the last step, the ratio to $y/L$ is bounded by a constant times > $L^{4-A+o(1)}L_3^2/L_2^4$, which tends to zero. > This discharges the smooth-number range rather than assuming that a > fixed-$u$ estimate applies while $u$ grows. ### O4. OPEN — Dyadic prime-count asymptotic (original Eq. 10) The coefficient-$1/2$ asymptotic is not proved. The literal inclusive reserved interval has the sufficient eventual lower bound $y/(3\log y)$ proved in Section 7. Original statement and domain (`original-manuscript-verbatim.md`, lines 186–192): > **Input P.** The prime number theorem gives > $$ > \pi(y)-\pi(y/2)\sim\frac{y}{2\log y}. > \tag{10} > $$ > No explicit prime-count threshold is required. This is the final counting > input in [KK, section 2] and [R4, section 7]. ### O5. OPEN — External one-class comparison and the general quoted sieve (original Eqs. 5–7, 11) The full external Ford–Green–Konyagin–Maynard–Tao one-class theorem and the quoted arbitrary-kappa/multiplicative-g/nonnegative-weight sieve are not established here as those statements. Actual incidence/CRT moments, the residue-set interpretation, and the required finite dimension-four sieve are proved. The finite cover identity is also proved, and is not an open obligation. Original statement and domain (`original-manuscript-verbatim.md`, lines 116–131): > The same argument with one class per prime gives the usual Jacobsthal > gap $g(P(y))$. Adding a second class cannot destroy a cover, so > $$ > G_2(P(y))\geq g(P(y)). > \tag{4} > $$ > Equation (4) is elementary. Inserting the published one-class lower bound > reported in [KK, Theorem A and reference 1] gives the inherited comparison > $$ > G_2(P(y))\gg > \frac{y\log y\log\log\log y}{\log\log y}. > \tag{5} > $$ > The cited one-class theorem, due to Ford, Green, Konyagin, Maynard and Tao, > is external mathematical input, not a result established by computation > in this assignment. Original statement and domain (`original-manuscript-verbatim.md`, lines 138–160): > **Input S ([KK, Lemma 1]).** Let $\kappa$ be fixed. For primes $p\leq z$, > suppose a multiplicative function $g$ satisfies > $0\leq g(p)\leq\kappa$ and $g(p) $w_t$ satisfy, for every squarefree $d\mid P(z)$, > $$ > \sum_{d\mid t}w_t=\frac{Xg(d)}d+r_d,\qquad |r_d|\leq g(d). > \tag{6} > $$ > With $z\ll X$, the quoted lemma bounds their sifted sum by > $$ > \sum_{\gcd(t,P(z))=1}w_t > \ll_\kappa X\prod_{p\leq z}\left(1-\frac{g(p)}p\right). > \tag{7} > $$ > We use only $z\leq X$; its proportionality constant is fixed. > The statement was read in arXiv v2, page 4, and in the journal's online > text. Its proof cites Halberstam and Richert, *Sieve Methods* (1974), > Theorem 2.2. That printed book theorem has not been independently read > here. A second carrier of the same statement is [FI], Theorem 6.9 with > Corollary 6.10: it was reached at OCR custody in return #158 and > re-read there, so it is not counted as a page reading. Sections 4--8 > establish deductions conditional on using (7) in this stated form. They > do not reconstruct its underlying sieve proof. Original statement and domain (`original-manuscript-verbatim.md`, lines 196–228): > **Lemma 2 (elementary reduction from Input S).** For sets > $\Omega_p\subset\mathbb Z/p\mathbb Z$ with > $g(p)=|\Omega_p|\leq\kappa$ and $g(p) $$ > S(X,\Omega)= > \#\{1\leq n\leq X:n\bmod p\notin\Omega_p > \text{ for every }p\leq z\}, > \qquad > V(z)=\prod_{p\leq z}\left(1-\frac{g(p)}p\right). > $$ > For integral $X$ and $z\ll X$, Input S implies > $$ > S(X,\Omega)\ll_\kappa X V(z). > \tag{11} > $$ > > **Proof.** For $1\leq n\leq X$, set > $$ > D(n)=\prod_{\substack{p\leq z\\ n\bmod p\in\Omega_p}}p, > \qquad > w_t=\#\{1\leq n\leq X:D(n)=t\}. > \tag{12} > $$ > The empty product is 1. These are nonnegative weights with finite support. > For squarefree $d\mid P(z)$, the condition $d\mid D(n)$ requires > $n\bmod p\in\Omega_p$ at every prime dividing $d$. By the Chinese > remainder theorem this is precisely $g(d)=\prod_{p\mid d}g(p)$ residue > classes modulo $d$. Counting an interval in each class gives (6), with > $|r_d|\leq g(d)$, for every such $d$, including $d>X$. > > Since $D(n)\mid P(z)$, it is coprime to $P(z)$ exactly when > $D(n)=1$. Hence the left side of (7) is exactly $S(X,\Omega)$. > Applying Input S proves (11). $\square$ ### O6. OPEN — Original harmonic-density expansion and two-sided Euler-product comparisons (Eqs. 20–21) The printed $O(1)$ harmonic-density expansion and both directions of the $\asymp$ product comparisons remain unproved as stated. One-sided finite inequalities (10)–(11) and the exact parameter substitution prove the required sieve budget. Original statement and domain (`original-manuscript-verbatim.md`, lines 355–387): > ### 6.2 The product calculation > > The two small primes contribute a bounded term, so Input M yields > $$ > \sum_{p\leq\sqrt y}\frac{g(p)}p > =2\log\log\sqrt y+ > 2(\log\log z_1-\log\log z_0)+O(1). > \tag{20} > $$ > The contribution of the middle band is additional to the basic > two-class contribution. Since $g(p)\leq4$ and the small primes are > handled separately, the quadratic and higher terms in the logarithm > of the product have bounded sum. It follows that > $$ > V(\sqrt y)\asymp > \frac1{(\log\sqrt y)^2} > \left(\frac{\log z_0}{\log z_1}\right)^2 > \asymp > \frac{A^4 L_2^4}{L^4L_3^2}. > \tag{21} > $$ > These are comparisons up to constant factors, not identities with unit > constant. In particular, exponentiating an $O(1)$ in (20) does not > remove it. > > Multiplying (21) by (13) gives > $$ > S(m,\Omega)\ll \frac{A^4}{B}\frac{y}{L}. > \tag{22} > $$ > After fixing $A$, choose $B$ large enough that this term is at most > $y/(4L)$ for all sufficiently large $y$. We have not assigned a > numerical value to the sieve constant. ### O7. OPEN — The separate short construction and its priced constant (original Eq. 28) The original $m=\lfloor c y\log y\rfloor$, $z=\sqrt m$ derivation, its $4Cc+o(1)$ coefficient and quoted numerical pricing remain unproved. The original Eq.27 endpoint is actually proved by PrimaryMain.eventual_y_log_gap through the main theorem; it is not classified as an unproved endpoint. Original statement and domain (`original-manuscript-verbatim.md`, lines 455–484): > For completeness, let $m=\lfloor c y\log y\rfloor$ and > $z=\sqrt m$, with $c>0$ fixed and sufficiently small. > Choose $a_p=0$ for $p\leq z$. > > Lemma 2 with $\kappa=2$, together with Mertens' theorem, bounds the > uncovered indices by > $$ > C\,\frac{m}{(\log z)^2} > \leq (4Cc+o(1))\,\frac{y}{\log y}. > \tag{28} > $$ > Here the constant $C$ absorbs the convergent local factors, including > the prime 2. Since $z=o(y)$, the number of unused primes is > $\pi(y)-\pi(z)\sim y/\log y$. Choosing $c<1/(8C)$ leaves more > than one unused prime per survivor, so the assignment in section 7 > finishes the cover. > > If that constant is priced at all, the cited form gives > $C=2C_2e^{-2\gamma}C_1=0.41621\,C_1$ with $C_1(1/2)=805.5$ > ($\delta=0.05$, $K=3$), so any $0 ([FI, Theorem 6.9 and Corollary 6.10], conditional on that reading). > The value $7.5\times10^{-4}$ recorded in return #158 is not reused: it > drops the local factor for the prime $p=2$ that $C$ carries. That is a > priced constant, not a certified threshold. > > This argument does not use the smooth-number input, but it still uses > Input S. It cannot be described as having escaped the unread > Halberstam--Richert source dependency. Its composition has the same > project-derived, unrefereed status. In asymptotic size, (1) is stronger > than (27), which is stronger than the inherited comparison (5). ## 10. Source correspondence and review boundary The source map accompanying this paper records each section's frozen files, declaration locators, exact source hashes, and statement qualifications. The mathematical chain is finite covering → actual incidence sieve → the smooth budget at $A=12$ and elementary prime supply → cover completion → integer maximum binding. Named intermediate estimates with hypotheses, such as the range conditions in (14), are used only after the parameter lemmas discharge those hypotheses. A definition of a target proposition is not treated as its proof. The proof package's axiom policy allows only Lean's standard `propext`, `Classical.choice`, and `Quot.sound`; individual closures may use a subset. There is no mathematical premise named Input S, M, H, or P in the three final theorem statements. This writing task inspected source and existing evidence and did not run a new proof build or independent checker. Compiler and kernel evidence belongs to its pinned proof package. The new manuscript hash requires an actual new statement-binding review before any public acceptance claim is made. The construction's mathematical lineage remains the original manuscript's project record and Kalmynin–Konyagin three-band architecture. In particular, their polynomial root problem and the fixed-separation pair here are different extremal objects; we do not identify their theorem directly with (1). The finite Möbius/diagonalization adaptation in `SelbergFinite.lean` retains its Arend Mellendijk copyright and PNT+ source credit. The ordinary primorial estimate and arithmetic-function identities are established library dependencies. Full uniform smooth asymptotics, prime-number-theorem asymptotics, and the general external sieve statement remain in Section 9. The adapted-source credits are explicit in the frozen Lean headers and remain in those files: - `SelbergFinite.lean` and `SelbergError.lean` adapt the finite Möbius/diagonalization and lambda-square/lcm-pair arguments from Arend Mellendijk's PNT+, revision `2667e414c38e5a5dc9aa1946f16f13001e5cd3ed`. Both preserve the 2023 copyright, author credit and Apache 2.0 notice. - `PrimeLog.lean`, `MertensBand.lean` and `SelbergEuler.lean` adapt OpenAI math (`openai/math`), revision `adc7f1241b42e322a6451854ab7e4b4c146bf78a`, under Apache 2.0. The headers identify the finite weighted prime-log, Mertens-factor/Rankin, finite Euler truncation and sieve-denominator arguments and the local modifications; no foreign project or toolchain is imported. - `PrimeLog.lean` also attributes its elementary theta fragments to Mathlib's `Chebyshev.lean`: copyright 2025 Alastair Irving; the header names Alastair Irving, Terry Tao and Ruben Van de Velde. The broader pinned Mathlib library retains its contributors' notices and attribution. Replacing an integral/harmonic argument locally does not remove the adapted source credit. These code notices and licenses are distinct from the project manuscript's prose publication terms. Attribution additions do not assert sole authorship, a new grant in external material, or completion of the separate source-ledger repair. [Job #3880](https://solveathome.org/projects/twin-primes/job/3880) remains a generated-index/source-provenance repair awaiting the specifically chosen source expert and authoritative history; this manuscript correction neither fabricates that chronology nor closes the handoff. Bibliographic provenance is preserved from the original rather than newly audited here: A. B. Kalmynin and S. V. Konyagin, *A polynomial analogue of Jacobsthal function*, *Izvestiya: Mathematics* **88** (2) (2024), 225–235, DOI 10.4213/im9467e, arXiv:2302.00459v2; A. Hildebrand and G. Tenenbaum, *Integers without large prime factors*, *Journal de Théorie des Nombres de Bordeaux* **5** (1993), 411–484; and the historical sieve citations to Halberstam–Richert and Friedlander–Iwaniec. The original reference list, source-access qualifications, and project contributions remain available verbatim. No novelty or independent literature-completeness claim is made by this rewrite.