\section{The construction}\label{sec:construction} \subsection{Classes and free data} Fix $0 < \eta < \alpha' < 1/2$. A component $u$ of a vector is \emph{small} if $\eta \le u \le \alpha'$, \emph{rough} if $\alpha' < u < 1/2$, and \emph{large} if $u \ge 1/2$. Fix a partition of $(\alpha', 1/2)$ into finitely many intervals, the \emph{bins}. A free vector $u$ has a \emph{class} $(j, k, b)$: $j$ small components, $k$ rough components, and $b$ the bin of its largest rough component ($b = 0$ if $k = 0$). Its \emph{small total} is the sum $\tau$ of its small components. The free data have the form \begin{equation}\label{eq:free} f(u) = c[j, k, b] + \mathbf 1[1 \le j \le m^*,\ \tau \le \alpha']\; w_k\, h_j(\tau), \end{equation} with finitely many constants $c[j, k, b]$, a \emph{profile} of weights $w_k$, and one function $h_j$ on $[j\eta, \alpha']$ for each level $1 \le j \le m^*$. The rungs use these profiles: \begin{center} \begin{tabular}{lll} \toprule rung & levels & profile \\ \midrule P3 & $m^* = 1$ & $w_k = 1$ for every $k$ \\ P4 & $m^* = 2$ & $w_k = 1$ for every $k$ \\ P5 & $m^* = 3$ & $w_3 = 1$, $w_k = 0$ for $k \ne 3$ \\ P6, P7 & $m^* = 7$ & $w_2 = 1$, $w_3 = -1$, $w_k = 0$ otherwise \\ \bottomrule \end{tabular} \end{center} The values on all other vectors are given by FM's extension (6.3). (For P4 the level-$2$ function is a function of the pair's total, as written; the masked functions of \S\ref{sec:lift} depend on the pair itself.) \subsection{The census} An $\eta$-rough squarefree modulus $d \le x^\alpha$ with $\alpha < \alpha'$ is $1$ or a product of $t$ primes, each at least $(x/2)^\eta$. Then $(x/2)^{t\eta} \le d \le x^\alpha$, so $t\eta \le \alpha/(1 - \log 2/\log x)$, which is less than $\alpha'$ for large $x$. We take $m^* = \max\{j : j\eta < \alpha'\}$. Then a modulus with $m^* + 1$ or more prime factors has $d \ge (x/2)^{(m^* + 1)\eta} \ge (x/2)^{\alpha'} > x^\alpha$ for large $x$, so only $t \le m^*$ occurs. When $(m^* + 1)\eta = \alpha'$ exactly (P4: $3 \cdot \tfrac1{18} = \tfrac16$; P5: $4 \cdot \tfrac1{20} = \tfrac15$; P7: $8 \cdot \tfrac1{30} = \tfrac4{15}$), a function $h_{m^*+1}$ could live only at the single point $\tau = \alpha'$, and we set it to zero: the \emph{null level}. The census then has no slack and uses $\alpha < \alpha'$ strictly. \subsection{The rows and the forcing} Unmasked, the free data \eqref{eq:free} see the small part of a vector only through $j$ and $\tau$. So the twisted functional at a tuple of $j$ small components with total $\sigma$ depends only on $\sigma$, and it has the form \begin{equation}\label{eq:row} T_j(\sigma) = A_j(\sigma)\cdot c + B_j(\sigma)\, h_j(\sigma) + \sum_{i \ge 1,\ j + i \le m^*} C_{j,i}[h_{j+i}](\sigma), \qquad \sigma \in [j\eta, \alpha'], \end{equation} where $A_j(\sigma)\cdot c$ collects the constants, the \emph{diagonal} $B_j(\sigma)$ collects the completions of the tuple that add no small component, weighted by $w_k$, and $C_{j,i}$ collects those that add exactly $i$ small components. No $h_J$ with $J < j$ enters, since the row's vectors have at least $j$ small components. If $B_j$ does not vanish on $[j\eta, \alpha']$, then \[ h_j := -\bigl(A_j \cdot c + \textstyle\sum_{i} C_{j,i}[h_{j+i}]\bigr)/B_j, \qquad j = m^*, m^* - 1, \dots, 1, \] defines the functions top-down and makes every row $T_j \equiv 0$ hold \emph{exactly}; each $h_j$ is linear in $c$. This is the \emph{forcing}. What remains to choose are the finitely many constants. The kernels and the rows in full are in Appendix~\ref{app:kernels}. For the profiles above the diagonal has a sign. For P5, $B_j$ uses only completions with exactly three rough parts, and the kernels of Appendix~\ref{app:kernels} give $B_j(\sigma) = 2(-1)^{j+1}(R_1 * R_2\mathbf 1_{[1/2,1]})(1-\sigma)$, which is nonzero exactly when $1-\sigma > 1/2 + \alpha'$, hence on all of $[j\eta,\alpha']$ when $\alpha' < 1/4$; at $\alpha' = 1/5$ its minimum is at $\sigma = \alpha'$ (value $0.8250801\ldots$), and the certificate's computation, as recorded by the audit's replay of it (\S\ref{sec:computations}), gives $\min_j|B_j| \ge 0.825079$. For the weighted profile there is also a closed form. \begin{lemma}[The weighted diagonal]\label{lem:Beff} Let $\alpha' \ge 1/4$, and let $R_k = r^{*k}/k!$, where $r(u) = u^{-1}$ on $(\alpha', 1/2)$ and $r(u) = 0$ elsewhere. For the profile $w_2 = 1$, $w_3 = -1$, \[ B_j(\sigma) = (-1)^j\bigl[2R_2(t) + 6R_3(t)\bigr], \qquad t = 1 - \sigma . \] In particular $(-1)^j B_j \ge 0$. If $\alpha' < 1/3$, then $B_j$ has the sign $(-1)^j$ and does not vanish on $[j\eta, \alpha']$; if $\alpha' \ge 1/3$, then $B_j \equiv 0$ on the nonempty interval $[\max(j\eta, 1 - 2\alpha'), \alpha']$. \end{lemma} \begin{proof} A completion of the tuple that adds no small component either is free, with $k$ rough parts, or has one large part that fragments into rough parts. Rough parts exceed $\alpha' \ge 1/4$, so $R_2$ and $R_3$ vanish on $[0, 1/2)$, and $R_1\mathbf 1_{[1/2, \infty)} = 0$. Hence the Lemma~\ref{lem:K} kernels of a large part fragmenting into $k_y$ rough parts are $\Lambda^0_2 = -R_2$ and $\Lambda^0_3 = 2R_3$. With the row sign $(-1)^{\dim}$, the contributions are, for $k = 2$, $(-1)^j R_2 + (-1)^{1+j}\Lambda^0_2 = 2(-1)^j R_2$, and for $k = 3$, $(-1)^{j+1}R_3 + (-1)^{1+j}\Lambda^0_3 + (-1)^{j}(R_1 * \Lambda^0_2) = (-1)^{j+1}(1 + 2 + 3)R_3$, using $R_1 * R_2 = 3R_3$. Weighting by $w_2 = 1$ and $w_3 = -1$ gives the claim. Since $R_2(t) > 0$ for $t \in (2\alpha', 1)$, which contains $1 - \sigma$ for every $\sigma \le \alpha'$ when $\alpha' < 1/3$, the diagonal then does not vanish. If $\alpha' \ge 1/3$, then $R_3(t) = 0$ for $t < 1 \le 3\alpha'$, and $R_2(t) = 0$ for $t \le 2\alpha'$, that is, for $\sigma \ge 1 - 2\alpha'$; for $\alpha' \ge 1/3$ these are the rows of Lemma~\ref{rem:dead}. \end{proof} \subsection{The parity solve and the target inequalities} The row $T_0$ is linear in $c$, with the $h_j$ entering through $c$. Choose a direction $g$ supported on classes with $j = 0$ (no small component). Then $A_j \cdot g = 0$ for every $j \ge 1$, since rows with a small prime meet only classes with $j \ge 1$, so moving along $c + t g$ leaves every $h_j$ unchanged, and the solve $T_0(c + t g) = 0$ is affine in $t$. The directions used are listed in Table~\ref{tab:rungs}: in the cases with $m^* \le 3$, $g$ is the unit vector of one class with no small and three rough components, and $g = e_{(0,3,2)} + e_{(0,3,3)}$ in the two cases with the weighted profile. With $c^* = c + t^* g$, the certificate (\S\ref{sec:certificates}) proves: \begin{equation}\label{eq:target} f(1) < 0 \qquad\text{and}\qquad f(u) \ge (1 - m)\, f(1)\ \text{ for every non-prime } \eta\text{-rough vector } u \ne E, \end{equation} for an explicit $m > 0$, including the vectors with one large component, whose values come from (6.3). In the two cases with $m^* = 1$ the certified statement is narrower: it covers every such vector with no subsum equal to $1/2$. No lower bound is asserted at $E$, nor in those two cases on the other vectors with a subsum equal to $1/2$; the mask of \S\ref{sec:lift} makes the masked extension vanish there; Theorem~\ref{thm:lift} uses \eqref{eq:target} only at vectors with no subsum equal to $1/2$. \subsection{The constants} The constants are found by a linear programme: minimise $\sum|c|$ subject to an upper bound on $f(1)$, to the positivity rows at grid points, to $T_0 = 0$ and to box constraints, on an exact oracle at a moderate mesh. This is provenance only. The certificate takes the constants as exact rationals, the binary64 values of the stored solution, and does not use the programme.