\section{Consequences for Goldbach}\label{sec:goldbach} Let $N$ be a large even integer and $L = \log N$. The Goldbach count is $G(N) = \sum_{m < N} \Lambda(m)\Lambda(N - m)$, and the conjectured size is $G(N) \sim \mathfrak S(N) N$, where \[ \mathfrak S(N) = 2C_2 \prod_{p \mid N,\ p > 2} \frac{p - 1}{p - 2}, \qquad C_2 = \prod_{p > 2}\Bigl(1 - \frac{1}{(p - 1)^2}\Bigr). \] Read as a sum over $m$, $G(N)$ pairs the von Mangoldt function with the sequence $m \mapsto \Lambda(N - m)$. The constructions of the previous sections reflect to this setting. We replace $\Lambda(N - m)$ by a nonnegative sequence that agrees with a local model of $\Lambda(N - m)$ in a list of statistics, but vanishes on all large primes $m$. The listed statistics, with Hooley's twisted condition at the levels of Theorem~\ref{thm:A}, therefore do not by themselves force a Goldbach-scale count. \subsection{The local model} Put $S = \mathfrak S(N)$, $w = \exp(\sqrt L)$ and $P_w = \prod_{p \le w,\ p \nmid 2N} p$, and define, for $1 \le m < N$, \[ b'_m = K_w\, \mathbf 1_{(m, N) = 1}\, \mathbf 1_{(N - m, P_w) = 1} \prod_{p \mid m,\ p > w} \frac{p - 1}{p - 2}, \qquad K_w = S \prod_{p \le w,\ p \nmid 2N} \frac{p - 1}{p - 2} \sim e^{\gamma}\sqrt L . \] This is a presieved Euler model of $\Lambda(N - m)$: its masses in intervals and in residue classes to small moduli are those of $\Lambda(N - m)$. \begin{lemma}[Class masses of $b'$]\label{lem:classmass} Fix $A, C_0 > 0$. Uniformly for intervals $I \subseteq [1, N)$, integers $1 \le q \le L^{C_0}$ and residues $r$, \[ \sum_{m \in I,\ m \equiv r \bmod q} b'_m = \frac{|I|}{\varphi(q)}\, \mathbf 1_{(N - r, q) = 1} + O_{A, C_0}(N L^{-A}). \] \end{lemma} For $\Lambda(N - m)$ in place of $b'_m$ the same estimate is the Siegel--Walfisz theorem; in the classes with $(N - r, q) > 1$ only powers of primes dividing $q$ occur, and they contribute $O(L^2)$. We use upper and lower sieve weights of dimension~$1$ in the form of the fundamental lemma of sieve theory \cite[Lemma~6.3]{IK}, \cite[\S6.5]{FIOpera}. Let $D = \exp(L^{3/4})$. There are real numbers $\lambda_e^\pm$, chosen once from $D$ and $w$, with $\lambda_1^\pm=1$, $|\lambda_e^\pm|\le1$ and $\lambda_e^\pm=0$ for $e\ge D$, such that $\sum_{e\mid n}\lambda_e^-\le\mathbf1_{n=1}\le\sum_{e\mid n}\lambda_e^+$ for every squarefree $n$ whose prime factors are at most $w$, and such that, for large $N$, for every set $\mathcal P$ of primes $p \le w$ and every multiplicative function $g$ with $0 \le g(p) < 1$ that satisfies the dimension condition $\prod_{y_1 \le p < y_2,\ p \in \mathcal P}(1 - g(p))^{-1} \le (\log y_2/\log y_1)(1 + K/\log y_1)$ for all $2 \le y_1 < y_2$, \begin{equation}\label{eq:FL} \sum_{e \mid P} \lambda^\pm_e\, g(e) = \bigl(1 + O_K(e^{-L^{1/4}})\bigr) \prod_{p \in \mathcal P}\bigl(1 - g(p)\bigr), \qquad P = \prod_{p \in \mathcal P} p, \end{equation} since $\log D/\log w = L^{1/4}$. We use $g(p) = 1/p$, and $g(p) = 1/(p - 1)$ on sets of odd primes; both satisfy the dimension condition with an absolute $K$. For a product $P$ of primes $p \le w$ put $\theta^\pm_P(n) = \sum_{e \mid (n, P)} \lambda^\pm_e$, so that $\theta^-_P(n) \le \mathbf 1_{(n, P) = 1} \le \theta^+_P(n)$. \begin{proof}[Proof of Lemma~\ref{lem:classmass}] Let $P(w) = \prod_{p \le w} p$. Every prime $p \le w$ with $p \mid 2N$ divides $N$, since $N$ is even, and for such $p$ the conditions $p \nmid m$ and $p \nmid N - m$ coincide. Hence $\mathbf 1_{(m, N) = 1}\mathbf 1_{(N - m, P_w) = 1} = \mathbf 1_{(N - m, P(w)) = 1}$ unless $m$ has a prime factor $p \mid N$ with $p > w$. At most $L/\log w = \sqrt L$ primes $p > w$ divide $N$, so at most $\sqrt L\, N/w$ integers $m < N$ are of this kind. Every $m < N$ has at most $\sqrt L$ prime factors $p > w$, so the last product in $b'_m$ is $1 + O(\sqrt L/w)$. As $K_w \ll \sqrt L$, \[ \sum_{m \in I,\ m \equiv r \bmod q} b'_m = K_w M + O(NL/w), \qquad M = \#\{n \in N - I : n \equiv N - r \bmod q,\ (n, P(w)) = 1\}. \] Every prime factor of $q$ is at most $L^{C_0} < w$, so $M = 0$ if $(N - r, q) > 1$. Otherwise the condition $(n, P(w)) = 1$ only concerns the primes $p \le w$ with $p \nmid q$; let $P$ be their product. By the Chinese remainder theorem an interval of length $|I|$ contains $|I|/(qe) + O(1)$ integers $n \equiv N - r \pmod q$ with $e \mid n$, for every $e \mid P$. So $M$ lies between the two numbers \[ \frac{|I|}{q} \sum_{e \mid P} \frac{\lambda^\pm_e}{e} + O(D) = \frac{|I|}{q} \prod_{\substack{p \le w\\ p \nmid q}} \Bigl(1 - \frac1p\Bigr) \bigl(1 + O(e^{-L^{1/4}})\bigr) + O(D), \] by \eqref{eq:FL} with $g(p) = 1/p$. The product is $\frac{q}{\varphi(q)} V(w)$, where $V(w) = \prod_{p \le w}(1 - 1/p)$. Since $2(1 - \frac12) = 1$ and $\bigl(1 - \frac{1}{(p - 1)^2}\bigr)\frac{p - 1}{p - 2}\bigl(1 - \frac1p\bigr) = 1$ for $p > 2$, the factors at the primes $p \le w$ cancel, and \[ K_w V(w) = \prod_{p > w}\Bigl(1 - \frac{1}{(p - 1)^2}\Bigr) \prod_{p \mid N,\ p > w} \frac{p - 1}{p - 2} = 1 + O(\sqrt L/w). \] Hence $K_w M = \frac{|I|}{\varphi(q)}\mathbf 1_{(N - r, q) = 1} + O(N e^{-L^{1/4}} + \sqrt L\, e^{L^{3/4}})$, and every error term is $O_A(N L^{-A})$. \end{proof} If every prime factor of $m$ exceeds some $y \ge w$, the last product in $b'_m$ is $1 + O(L/(y \log y))$. So on integers without small prime factors, $b'$ is $K_w(1 + O(L/(y\log y)))$ times the constant baseline $b = 1$ of \S\ref{sec:setting}, times $\mathbf 1_{(m, N) = 1}$ and a presieve condition on $N - m$ with small moduli. The proof below transfers the estimates of \S\ref{sec:lift} through that presieve. \subsection{The reflected sequences} \begin{proposition}\label{prop:G} Let $(\eta, \alpha')$ be one of the constructions of Table~\ref{tab:goldbach} below. Fix $0 \le \alpha < \alpha'$, $0 < \nu < \eta$ and $C > 0$. For every large even $N$ there is a sequence $a_m \ge 0$ on $[1, N)$, depending only on the construction, $\alpha$, $\nu$, $C$ and $N$, that has the following properties for all $A, B > 0$. Implied constants depend only on the construction and on $\alpha$, $\nu$, $A$, $B$ and $C$. \begin{enumerate} \item[(G0)] $a_m = 0$ for every prime $m \ge N/L^C$. Moreover, $\sum_m a_m \Lambda(m) \ll S N L^{-C} + N^{1/2} L$. \item[(G1)] (Type I at the Bombieri--Vinogradov level, with interval maxima) \[ \sum_{d \le N^{1/2} L^{-C/2}} \tau(d)^{B} \max_{I} \Bigl|\sum_{m \in I,\ d \mid m} (a_m - b'_m)\Bigr| \ll N L^{-A}, \] the maximum over intervals $I \subseteq [1, N)$. \item[(G2)] (Type II on small factors) For all real coefficients $\xi_\ell$ and $\kappa_k$, \[ \sum_{1 < \ell \le N^\nu}\ \sum_{k < N/\ell} \xi_\ell\, \kappa_k\, (a - b')_{\ell k} = 0 . \] \item[(G3)] (Hooley's condition at level $\alpha$) \[ \sum_{q \le N^\alpha} \Bigl|\sum_{m < N,\ q \mid m} \mu(m)(a_m - b'_m)\Bigr| \ll N L^{-A}. \] \item[(G4)] (Parity, on every interval) For $f \in \{\mu, \lambda\}$, uniformly in intervals $I \subseteq [1, N)$, \[ \sum_{m \in I} (a_m - b'_m) f(m) \ll N L^{-A}. \] \item[(G5)] (Roughness comparisons) For every $z \le (N/(2L^C))^{\eta}$, the sums in (G4) restricted to the $z$-rough $m$, those all of whose prime factors are at least $z$, equal the unrestricted ones. \end{enumerate} \end{proposition} For even $N$, $S \ge 2C_2 > 1$, so $N^{1/2} L \le S N L^{-C}$ for large $N$, and (G0) gives $\sum_m a_m \Lambda(m) \ll S N L^{-C}$. This contradicts, for large $N$, every lower bound of the form $\sum_m a_m \Lambda(m) \ge \eps\, \mathfrak S(N) N$ with fixed $\eps > 0$. Hence no argument that proves such a lower bound for every nonnegative sequence satisfying (G1)--(G5) relative to $b'$ exists. Since $\Lambda(N - m)$ satisfies neither (G2) nor (G5) as stated (\S\ref{ssec:goldbach-scope}), the conclusion applies to arguments that use $\Lambda(N - m)$ only through (G1), (G3) and (G4), through bilinear estimates over factors in $(1, N^\nu]$, measured against $b'$, which (G2) implies, and through bounds for the sums of (G4) restricted to the $z$-rough $m$, $z \le (N/(2L^C))^\eta$, which (G4) and (G5) imply. \emph{Construction.} Fix $\alpha_*$ and $\nu_*$ with $\alpha < \alpha_* < \alpha'$ and $\nu < \nu_* < \eta$, and let $\eps > 0$ and $F_\eps$ be given by Theorem~\ref{thm:lift} for the construction and the pair $(\nu_*, \alpha_*)$; neither depends on $N$. Put $T = N/L^C$, $x_j = N/2^j$ for $0 \le j \le J = \lfloor C\log_2 L\rfloor$, and $B_j = (x_j/2, x_j] \cap [T, N)$. These sets partition $[T, N)$. The last one, $B_J = [T, x_J]$, is in general a proper subinterval of $(x_J/2, x_J]$, since $x_J/2 < T \le x_J$. Put \[ a_m = b'_m\,\bigl(1 + F_\eps(v(m))\bigr) \quad (T \le m < N), \qquad a_m = b'_m \quad (1 \le m < T), \] where $F_\eps(v(m))$ is read as $0$ unless $m$ is squarefree and $\eta$-rough. Then $a \ge 0$, because $F_\eps \ge -1$. For large $N$, $N^\alpha \le x_j^{\alpha_*}$ and $N^\nu \le x_j^{\nu_*}$ in every block, since $x_j \ge N/L^C$. Put $R = (N/(2L^C))^\eta$. For large $N$, $N^\nu < R$ and $R > w$, and the support of $a - b'$ consists of squarefree $m \ge T$ all of whose prime factors exceed $R$: if $m \in B_j$, its prime factors are at least $m^\eta > (x_j/2)^\eta \ge R$. The only new analytic step is the passage from the baseline $b = 1$ of Theorem~\ref{thm:lift} to the presieved weight $b'$. \begin{lemma}[Transfer through the presieve]\label{lem:transfer} Let $B_j$ be a block of the construction and $x = x_j$, and let $\mathcal Q$ be the set of squarefree $q \le x^{1/2}$ whose prime factors all exceed $R$ (so $1 \in \mathcal Q$). For $q \in \mathcal Q$, an interval $I \subseteq B_j$ and $f \in \{1, \mu\}$ put \[ \Sigma_f(q, I) = \sum_{\substack{m \in I\\ q \mid m}} f(m)\,(a_m - b'_m) = \sum_{\substack{m \in I\\ q \mid m}} f(m)\, b'_m\, F_\eps(v(m)), \qquad \Sigma^\flat_f(q, I) = \sum_{\substack{m \in I\\ q \mid m}} f(m)\, F_\eps(v(m)). \] There is a number $c_N = \mathfrak S(N)\bigl(1 + O(e^{-L^{1/4}})\bigr)$, depending only on $N$, such that for all fixed $A, B > 0$ and uniformly in the block \[ \sum_{q \in \mathcal Q} \tau(q)^B \max_{I \subseteq B_j} \bigl|\Sigma_f(q, I) - c_N\, \Sigma^\flat_f(q, I)\bigr| \ll x L^{-A}. \] \end{lemma} \begin{proof} Let $\mathcal S$ be the set of squarefree $\eta$-rough $m \in B_j$, and $C' = \sup |F_\eps|$. Every $m \in \mathcal S$ has at most $1/\eta$ prime factors, all exceeding $R > w$, and for $q \in \mathcal Q$ at most $x/q$ elements of $\mathcal S$ are multiples of $q$. We use $\sum_{q \le x} \tau(q)^B/q \ll L^{2^B}$, $K_w \ll \sqrt L$ and $\mathfrak S(N) \ll \log L$ throughout. \emph{Step 1: the Euler factor and $(m, N) = 1$.} On $\mathcal S$, $\prod_{p \mid m,\ p > w}\frac{p - 1}{p - 2} = 1 + O_\eta(R^{-1})$. At most $2/\eta$ primes $p > R$ divide $N$, and on $\mathcal S$ we have $\mathbf 1_{(m, N) = 1} = 1$ unless $m$ is a multiple of one of them. So, uniformly in $I$, with an additive error, \[ \Sigma_f(q, I) = K_w\, \Sigma^+_f(q, I) + O\Bigl(\frac{K_w x}{qR} + K_w \sum_{p \mid N,\ p > R} \Bigl(\frac{x}{[q, p]} + 1\Bigr)\Bigr), \] where $\Sigma^+_f(q, I) = \sum_{m \in I,\ q \mid m} f(m)\, F_\eps(v(m))\, \mathbf 1_{(N - m, P_w) = 1}$. Weighted by $\tau(q)^B$ and summed over $\mathcal Q$, the error is $\ll x L^{O_B(1)} R^{-1} + x^{1/2} L^{O_B(1)} \ll x N^{-\eta/2}$. \emph{Step 2: the presieve.} With the weights of \eqref{eq:FL} for the primes dividing $P_w$, write $\theta^\pm = \theta^\pm_{P_w}$. Since $0 \le \theta^+ - \mathbf 1_{(\cdot, P_w) = 1} \le \theta^+ - \theta^-$ and $|F_\eps| \le C'$, \[ \Sigma^+_f(q, I) = \sum_{e \mid P_w} \lambda^+_e\, \Sigma^\flat_f(q, I; e) + O\Bigl(C' \sum_{e \mid P_w} (\lambda^+_e - \lambda^-_e)\, \#\{m \in \mathcal S : q \mid m,\ m \equiv N \bmod e\}\Bigr), \] where $\Sigma^\flat_f(q, I; e)$ is $\Sigma^\flat_f(q, I)$ restricted to $m \equiv N \pmod e$. As $e \mid P_w$ is coprime to $2N$ and to every element of $\mathcal S$, this is a reduced residue class. \emph{Step 3: the Bombieri--Vinogradov theorem in the largest prime.} Let $\psi$ be either $m \mapsto f(m)\, F_\eps(v(m))$ or the indicator function of $\mathcal S$. We show that for every fixed $A_1$ \begin{equation}\label{eq:transfer-bv} \sum_{q \in \mathcal Q} \tau(q)^B \sum_{\substack{e \mid P_w\\ e < D}} \max_{I \subseteq B_j} \Bigl|\sum_{\substack{m \in I,\ q \mid m\\ m \equiv N \bmod e}} \psi(m) - \frac{1}{\varphi(e)} \sum_{\substack{m \in I\\ q \mid m}} \psi(m)\Bigr| \ll x L^{-A_1}. \end{equation} Let $m \in \mathcal S$ with $q \mid m$. As $m > x/2 > q$, the cofactor $m/q$ exceeds $1$; write $m = qhP$ with $P$ the largest prime factor of $m/q$. This factorisation is unique, $P > R$, and $m \equiv N \pmod e$ if and only if $P \equiv N\overline{qh} \pmod e$, again a reduced class. Fix $q$, $h$ and a dyadic range $Y < P \le 2Y$. For real $t \in (Y, 2Y]$ let $v_t$ be the increasingly ordered vector of the numbers $\log p/\log(qht)$, $p \mid qh$, and $\log t/\log(qht)$. Between the $O_\eta(1)$ values of $t$ at which $t$ passes a prime factor of $q$, the order of the coordinates is fixed and $v_t$ moves monotonically along a line segment (its coordinates are affine in $1/\log(qht)$), by $O(1/\log x)$ in all. By Lemma~\ref{lem:L2}, $F_\eps$ is in each dimension a finite sum of functions supported and Lipschitz on convex polytopes, and the conditions $qht \in I$, $t > P^+(h)$ (the largest prime factor of $h$) and the $\eta$-roughness of $qht$ cut out intervals of $t$. The sign $f(qhP)$ does not depend on $P$. So $t \mapsto \psi(qht)$, with the squarefree condition dropped, has total variation $O(1)$ on $(Y, 2Y]$, uniformly in $q$, $h$ and $I$, and partial summation bounds the difference between the sum of $\psi(qhP)$ over the primes $P \equiv c \pmod e$ in the range and $1/\varphi(e)$ times its sum over all primes in the range by \[ \ll \max_{y \le 2Y} \Bigl|\pi(y; e, c) - \frac{\pi(y)}{\varphi(e)}\Bigr|. \] The $O(1)$ primes $P \mid q$, excluded by the squarefree condition, change this by $O(1)$ for each $e$, hence by $O(D)$ after summing over $e$. By the Bombieri--Vinogradov theorem \cite{Bombieri65}, \cite{Vinogradov65}, \cite[Chapter~17]{IK}, in the form for $\pi(y; e, c)$, which follows from the form for $\psi(y; e, c)$ by partial summation, for every $A_2$ there is $B_2$ with \[ \sum_{e \le Y^{1/2}(\log Y)^{-B_2}} \max_{(c, e) = 1} \max_{y \le 2Y} \Bigl|\pi(y; e, c) - \frac{\pi(y)}{\varphi(e)}\Bigr| \ll_{A_2} Y (\log Y)^{-A_2}, \] and $D \le Y^{1/2}(\log Y)^{-B_2}$ for $Y \ge R/2$ and large $N$. For fixed $q$ and $Y$ there are at most $x/(qY)$ values of $h$. Summing over $h$, over the $O(L)$ dyadic ranges and over $q \in \mathcal Q$ with the weight $\tau(q)^B$ gives \eqref{eq:transfer-bv} with $A_1 = A_2 - 1 - 2^B$, plus $O(D\, x L^{O(1)}/R)$ from the excluded primes, which is $\ll x N^{-\eta/2}$. \emph{Step 4: assembly.} The density $1/(p - 1)$ at the primes $p \mid P_w$, all of which are at least $3$, satisfies the dimension condition, so by \eqref{eq:FL} $\sum_e \lambda^\pm_e/\varphi(e) = \bigl(1 + O(e^{-L^{1/4}})\bigr)\prod_{p \mid P_w}\bigl(1 - \frac{1}{p - 1}\bigr)$, and $K_w \prod_{p \mid P_w}\bigl(1 - \frac{1}{p - 1}\bigr) = \mathfrak S(N)$ by the definition of $K_w$. Put $c_N = K_w \sum_e \lambda^+_e/\varphi(e)$. Inserting \eqref{eq:transfer-bv}, for both choices of $\psi$, into Step~2 gives, uniformly in $I$, \[ K_w\, \Sigma^+_f(q, I) = c_N\, \Sigma^\flat_f(q, I) + O\Bigl(\mathfrak S(N)\, e^{-L^{1/4}}\, \frac{x}{q}\Bigr) + K_w\, \mathcal E_q, \qquad \sum_{q \in \mathcal Q} \tau(q)^B\, \mathcal E_q \ll x L^{-A_1}. \] With Step~1 and $A_1 = A + 1$, this proves the lemma. \end{proof} \begin{proof}[Proof of Proposition~\ref{prop:G}] We use the notation of Lemma~\ref{lem:transfer}, with $x = x_j$ in the block $B_j$. Lemma~\ref{lem:L3} and its proof hold with $(x/2, x]$ replaced by any subinterval $(cx, c'x]$, $1/2 \le c \le c' \le 1$, because \cite[Lemma~5.11]{FM}, on which the proof rests, is uniform in $c$ and $c'$; the integral of Lemma~\ref{lem:L3} (in its variable $w$ and with its $\lambda = \lambda_1 + \dots + \lambda_t$; these are not the presieve level and the sieve weights of this section) then runs over $[1 - \lambda + \log c/\log x,\ 1 - \lambda + \log c'/\log x]$, and the integrand is still zero. The same holds for the case $d = 1$ treated in Step~4 of \S\ref{sec:lift}, with $\Li(c'x) - \Li(cx)$ in place of $\Li(x) - \Li(x/2)$. Every block, the truncated $B_J$ included, has the form $(cx_j, c'x_j] \cap \mathbb Z$ with $1/2 \le c \le c' \le 1$, and so has every interval contained in a block. \begin{itemize} \item \emph{(G0).} A prime $p$ is squarefree and $\eta$-rough with $v(p) = (1)$, and $F_\eps(1) = -1$, so $a_p = 0$ for primes $p \ge T$. On proper prime powers $F_\eps(v(m))$ is read as $0$, so there $a_m = b'_m \le K_w(1 + O(1/w)) \ll \sqrt L$; as $\sum_{p^k < N,\ k \ge 2}\log p \ll N^{1/2}$, they contribute $\ll N^{1/2} L$. For the primes $w < p < T$, $a_p = b'_p \le K_w(1 + O(1/w))\mathbf 1_{(N - p, P_w) = 1}$. With the weights $\lambda^+_e$ and the Bombieri--Vinogradov theorem at level $D \le T^{1/2} L^{-B_2}$, \begin{align*} \#\{w < p < T : (N - p, P_w) = 1\} &\le \sum_{e \mid P_w} \lambda^+_e\, \pi(T; e, N) + O(wD) = \pi(T)\sum_{e \mid P_w} \frac{\lambda^+_e}{\varphi(e)} + O(T L^{-2})\\ &= \frac{\mathfrak S(N)}{K_w}\,\pi(T)\bigl(1 + o(1)\bigr) + O(T L^{-2}). \end{align*} The primes $p \le w$ contribute $O(K_w w \log w)$. Hence $\sum_{p < T} a_p \log p \ll \mathfrak S(N)\, N L^{-C}$, and (G0) follows. \item \emph{(G1).} Let $d \le N^{1/2}L^{-C/2}$ and let $I \subseteq [1, N)$ be an interval. The terms with $m < T$ vanish, so $\sum_{m \in I,\ d \mid m}(a_m - b'_m) = \sum_j \Sigma_1(d, I \cap B_j)$. This is $0$ unless $d$ is squarefree with all prime factors greater than $R$, and then $d \in \mathcal Q$ for every $j$, since $x_j \ge N/L^C$ gives $d \le x_j^{1/2}$. By Lemma~\ref{lem:transfer} with $f = 1$, and since $c_N \ll \log L$, it suffices to show $\sum_{d \le x_j^{1/2}} \tau(d)^B \max_I |\Sigma^\flat_1(d, I \cap B_j)| \ll x_j L^{-A - 1}$. Here $\Sigma^\flat_1(d, \cdot)$ is the Type~I sum of the sequence $w_n = F_\eps(v(n))$ of Theorem~\ref{thm:lift} at scale $x_j$, over the multiples of $d$ in an interval of $(x_j/2, x_j]$. Condition (I) of Theorem~\ref{thm:lift}, with $\max(A + 2, B)$ in place of its $B$, gives the bound, since $\log x_j \ge L/2$. Summing over the $J + 1 \ll \log L$ blocks gives (G1). \item \emph{(G2).} Every element of the support of $a - b'$ has all its prime factors greater than $R > N^\nu$. So no $\ell \in (1, N^\nu]$ divides an element of the support, and every term of the sum in (G2) vanishes. \item \emph{(G3).} Let $q \le N^\alpha$. As in (G1), $\sum_{m < N,\ q \mid m}\mu(m)(a_m - b'_m) = \sum_j \Sigma_\mu(q, B_j)$, and this is $0$ unless $q \in \mathcal Q$: a non-squarefree $q$ gives $0$ through $\mu$, and a $q$ with a prime factor at most $R$ divides no element of the support. Let $q = p_1 \cdots p_t \in \mathcal Q$. Then $t \le m^* = \max\{j : j\eta < \alpha'\}$: otherwise $t\eta \ge (m^* + 1)\eta \ge \alpha'$, and $q > R^t \ge (N/(2L^C))^{\alpha'} > N^\alpha$ for large $N$, because $\alpha < \alpha'$ strictly. (At $(1/18, 1/6)$, $(1/20, 1/5)$ and $(1/30, 4/15)$, $(m^* + 1)\eta = \alpha'$ exactly, so these cases use the strictness.) Also $q \le N^\alpha \le x_j^{\alpha_*} \le x_j^{1/2}$ in every block. By Lemma~\ref{lem:transfer} with $f = \mu$ and $I = B_j$ (the weight $\tau(q)^B \ge 1$ may be dropped), and since $c_N \ll \log L$, it suffices to show $\sum_q |\Sigma^\flat_\mu(q, B_j)| \ll x_j e^{-(\log x_j)^{1/5}}$, the sum over $q \in \mathcal Q$ with $q \le N^\alpha$. In the block of scale $x = x_j$, every element of the support has all prime factors at least $(x/2)^\eta$, so $\Sigma^\flat_\mu(q, B_j)$ is $0$ unless every $p_i \ge (x/2)^\eta$. In that case, for $t \ge 1$, $\Sigma^\flat_\mu(q, B_j)$ is the sum $S^\mu_q$ of Lemma~\ref{lem:L3} at scale $x$, with $\alpha_*$ in place of $\alpha$, restricted to $B_j$, and Lemma~\ref{lem:L3} in the subinterval form above gives $\ll (x/q)\, e^{-(\log x)^{1/5}}$, because $T_t(F_\eps)$ vanishes at every $t$-tuple with total at most $\alpha'$. For $q = 1$ the case $d = 1$ of Step~4 of \S\ref{sec:lift}, in the subinterval form, gives $(\Li(c'x) - \Li(cx))\, T_0(F_\eps) + O(x e^{-(\log x)^{1/5}})$, and $T_0(F_\eps) = 0$. Summing over $q$ costs at most a factor $\sum_{t \le m^*}\bigl(\sum_{R < p \le N} 1/p\bigr)^t \ll 1$, since $\sum_{R < p \le N} 1/p = \log(1/\eta) + o(1)$. Summing over the blocks gives (G3). \item \emph{(G4).} Let $I \subseteq [1, N)$ be an interval. As in (G1), $\sum_{m \in I}(a_m - b'_m)\mu(m) = \sum_j \Sigma_\mu(1, I \cap B_j)$. By Lemma~\ref{lem:transfer} with $q = 1$ and $f = \mu$ it suffices that $\Sigma^\flat_\mu(1, (cx_j, c'x_j]) \ll x_j e^{-(\log x_j)^{1/5}}$ uniformly for $1/2 \le c \le c' \le 1$. This is the subinterval form of the case $d = 1$: the sum is $(\Li(c'x_j) - \Li(cx_j))\, T_0(F_\eps) + O(x_j e^{-(\log x_j)^{1/5}})$, and $T_0(F_\eps) = 0$. So (G4) holds for $f = \mu$; and $a - b'$ is supported on squarefree integers, where $\lambda = \mu$. \item \emph{(G5).} The support of $a - b'$ consists of integers whose prime factors all exceed $R$. They are $z$-rough for every $z \le R$, so restricting the sums in (G4) to $z$-rough $m$ changes nothing. \qedhere \end{itemize} \end{proof} \begin{table}[ht] \centering \begin{tabular}{@{}lllc@{}} \toprule Construction $(\eta, \alpha')$ & Twisted level & Type II range & Separate implementations \\ \midrule $(1/11, 1/10)$ & $\alpha < 1/10$ & $\nu < 1/11$ & 1 \\ $(1/14, 1/8)$ & $\alpha < 1/8$ & $\nu < 1/14$ & 1 \\ $(1/18, 1/6)$ & $\alpha < 1/6$ & $\nu < 1/18$ & 2 \\ $(1/20, 1/5)$ & $\alpha < 1/5$ & $\nu < 1/20$ & 2 \\ $(1/30, 1/4)$ & $\alpha < 1/4$ & $\nu < 1/30$ & 1 \\ $(1/30, 4/15)$ & $\alpha < 4/15$ & $\nu < 1/30$ & 1 \\ \bottomrule \end{tabular} \caption{The cases of Proposition~\ref{prop:G}. Each uses the certified function of the stated construction and its lift (Theorem~\ref{thm:lift}). The last column is the number of separately written implementations that certify the construction (\S\ref{ssec:status}).}\label{tab:goldbach} \end{table} The base case $(1/11, 1/11)$ of Remark~\ref{rem:base} needs no row: its ranges $\alpha < 1/11$ and $\nu < 1/11$ lie inside those of $(1/11, 1/10)$, so it follows as a corollary of that case, without a separate certificate. The reflection argument for each case was audited independently, in an earlier written form: the audit reconstructed the passage from the certified function and its lift to (G0)--(G5), confirmed it, and required corrections, which the proof above incorporates; for $(1/20, 1/5)$ it also covered the cap computation below. These audits check the reflection, not the constructions. The constructions' certification is that of \S\ref{ssec:status}: $(1/18, 1/6)$ and $(1/20, 1/5)$ are certified by two separately written implementations each and the others by one; for $(1/30, 1/4)$ and $(1/30, 4/15)$ that one is the second-order engine, whose code was audited, and no second certificate exists. \subsection{What the proposition does and does not say}\label{ssec:goldbach-scope} \begin{itemize} \item \emph{Strict ranges only.} Each case holds for every fixed $\alpha$ below its level and every fixed $\nu$ below its $\eta$. No endpoint is claimed. No threshold beyond the last case is asserted, and none follows from Conjecture~\ref{conj:wall}, which concerns a family of constructions, not Goldbach's problem. \item \emph{Type II is stronger than the truth's.} (G2) is an exact identity for arbitrary coefficients, and the true sequence fails that form. Take a prime $\ell\le2L$ with $\ell\nmid N$. Such a prime exists for large $N$: otherwise $\prod_{p\le2L}p$ divides $N$, whereas its logarithm is $2L+o(L)>L$ by the prime number theorem. Set $\xi = \mathbf 1_{\{\ell\}}$, and $\kappa_k = \operatorname{sgn}(\Lambda(N - \ell k) - b'_{\ell k})$ for $N/2 < \ell k < N$, $\kappa_k = 0$ otherwise. The sum is then $\sum_{N/2 < \ell k < N} |\Lambda(N - \ell k) - b'_{\ell k}| \asymp N/\ell$, by the prime number theorem in progressions and Lemma~\ref{lem:classmass}, so the true sequence fails even Ford and Maynard's bounded-coefficient form (II) with $\theta = 0$ on the block $(N/2, N)$, for every $\nu > 0$. So (G2) is not an input that the truth provides. What it shows is that every weaker small-factor bilinear bound measured against $b'$ also holds for the reflected sequence, whatever coefficient class the true sequence satisfies, because the reflected sums vanish for all coefficients. \item \emph{Type I is shared only up to a common level.} (G1) reaches $d \le N^{1/2}L^{-C/2}$. For $\Lambda(N - m)$, the sums over the multiples of $d$ are controlled by the Bombieri--Vinogradov theorem (the weight $\tau(d)^B$ is removed by the Cauchy--Schwarz inequality; for $(d, N) > 1$ only powers of primes dividing $(d, N)$ occur, as after Lemma~\ref{lem:classmass}) only for $d \le N^{1/2}L^{-B'}$, with some $B' = B'(A, B)$. That $b'$ satisfies the matching estimate, $\sum_{d \le N^{1/2}} \tau(d)^B \max_I \bigl|\sum_{m \in I,\ d \mid m} b'_m - \mathbf 1_{(d, N) = 1}|I|/\varphi(d)\bigr| \ll N L^{-A}$, follows from the fundamental lemma as in the proof of Lemma~\ref{lem:classmass}, with the Euler factor of $b'_m$ expanded as a divisor sum; we do not write this out. The two sequences are therefore compared on $d \le \min(N^{1/2}L^{-C/2}, N^{1/2}L^{-B'})$; when $C < 2B'$ the longer range of the reflected sequence is an extra property, like (G2) and (G3). \item \emph{Hooley's condition and the parity sums are granted.} (G3) is not known for $\Lambda(N - m)$ at any fixed power level, and (G4) is not known for it at the precision $N L^{-A}$, even on $[1, N)$ with $f = \mu$. There the $b'$-part $\sum_{m < N} b'_m \mu(m)$ is $\ll N L^{-A}$ (by the presieve weights of \eqref{eq:FL}, with the Euler factor of $b'_m$ expanded as a divisor sum, and the Bombieri--Vinogradov theorem for $\mu$ to the moduli below $D$; we omit the details), so (G4) amounts to a bound for the M\"obius function on the shifted primes $N - p$. What is known for the true sequence is weaker and recent. L\"u, Wang and Wu give $\sum_{p \le x} \mu(p + 2) = o(\pi(x))$ \cite{LWW}. Guo deduces from their arguments $\sum_{p < N} \mu(N - p) = o(\pi(N))$, and the same for $\lambda$, with the rate $O(\pi(N)/\sqrt{\log\log N})$ \cite{Guo}. He assumes their arguments as written and, for general $N$, a zero-free region for Dirichlet $L$-functions; when $\operatorname{rad}(N) \le (\log N)^{O(1)}$ he does not need the second assumption. Both works are recent preprints. (G5) is, like (G2), an exact identity for the reflected sequence, and the true sequence fails it as written: for $f=\lambda$, let $2^k$ be the largest power of two strictly below $N$, set $m=N-2^k$, and take $I=\{m\}$ and $z=3\le R$. Then $00$. The unrestricted singleton sum is nonzero, while its restriction to $z$-rough integers is zero. Including (G3)--(G5) only strengthens the conclusion. \item \emph{Huang and Li's criterion is untouched.} Their twisted hypothesis concerns the true Goldbach sequence, subtracts the full initial-interval sum divided by $\varphi(q)$, and takes maxima over initial intervals and reduced residue classes. Its level is $N^{1-\theta}$, paired with prime Elliott--Halberstam at $N^\theta(\log N)^{2A+8}$, in their notation (\S\ref{sec:intro}). In the specialization where the latter is supplied by Bombieri--Vinogradov, $\theta<1/2$, so the twisted level exceeds $1/2$. Proposition~\ref{prop:G} grants (G3) only below $4/15$, measured against $b'$, and establishes no version of their criterion. \item \emph{Plain and switching versions.} The proposition asserts no explicit cap $a_m \le K b'_m$; the bound $|F_\eps| \le C'$ of \S\ref{sec:lift}, Step~3, gives one only with an inexplicit $K$. Explicit caps matter for Chen-type switching arguments. They need an upper enclosure of $f$, and the certificates of \S\ref{sec:certificates} bound $f$ on non-prime vectors only from below. For $(1/20, 1/5)$, an upper-bound pass added to the code of the certificate of record and run at its settings ($M = 7200$; its lower-bound outputs agree with the stored certificate field by field) gives, in exact rational arithmetic, $f(u) \le 2011282180584704051/(5\cdot 10^{17}) < 4.022565$ for every non-prime $\eta$-rough vector $u \ne E$. With $|f(1)| \ge 527670700557531721/(5\cdot 10^{17})$, this gives $F \le 2011282180584704051/527670700557531721 < 3.811624$. Since $f_\eps = f + O(\eps)$ off the mask and $f_\eps = 0$ on it (\S\ref{sec:lift}, Step~3), $F_\eps < 4$ for $\eps$ small enough, and then the reflected sequence satisfies $a_m \le 5b'_m$. The same computation gives $F > 3.8$ on an open region of two-component vectors. Take an open box $\mathcal V$ inside it whose closure contains no vector with a component $1/2$; for small $\eps$, $\mathcal V$ is unmasked and $F_\eps > 3$ on it. The reflected indices meet $\mathcal V$ with positive $b'$-mass for every large $N$: Steps~1--4 of the proof of Lemma~\ref{lem:transfer}, with $q = 1$, $f = 1$ and $F_\eps$ replaced by the indicator function of $\mathcal V$ (a convex polytope, so Step~3 applies unchanged), give $\sum b'_m = c_N\, \#\{m\} + O(N L^{-2})$ over the squarefree $m \in B_0$ with $v(m) \in \mathcal V$, and there are $\gg N/L$ such $m$. So $a_m \le 4b'_m$ fails for this function. The pass is one implementation; an independent audit read its code and reran it at $M = 3600$ and $7200$ (\S\ref{sec:computations}). No cap is claimed for the other cases. \end{itemize}