\section{Setting}\label{sec:setting} \subsection{Type I, Type II and Hooley's condition} We follow Ford and Maynard \cite{FM}. Let $(a_n)$ and $(b_n)$ be real sequences supported on $(x/2, x]$, with $a_n \ge 0$, and put $w_n = a_n - b_n$. For parameters $0 < \gamma < 1$, $0 \le \theta < 1/2$ and $0 < \nu \le 1 - \theta$, a constant $\varpi > 1$ and a large constant $B$, the hypotheses are: \begin{itemize} \item[(I)] (Type I range $[0, 1 - \gamma]$) $\displaystyle\sum_{m \le x^\gamma} \tau^B(m) \max_{I} \Bigl|\sum_{\substack{x/2 < mn \le x\\ n \in I}} w_{mn}\Bigr| \le \frac{x}{\log^B x}$, the maximum over intervals $I$; \item[(II)] (Type II range $[\theta, \theta + \nu]$) $\displaystyle\Bigl|\sum_{\substack{(x/2)^\theta < \ell \le x^{\theta+\nu}\\ x/2 < \ell n \le x}} \xi_\ell \kappa_n w_{\ell n}\Bigr| \le \frac{x}{\log^B x}$ for all $|\xi_\ell| \le \tau^B(\ell)$, $|\kappa_n| \le \tau^B(n)$; \item[(w)] (Ford and Maynard's mild growth condition \cite[\S2]{FM}) $\displaystyle\sum_n |w_n|\tau(n) \le x(\log x)^{\varpi}$ and $w_n \ge -x^{\nu/10}$ for $x/2 < n \le x$. \end{itemize} FM also require $b_n \ge 0$ and two conditions on the comparison sequence: a lower bound $\sum_p b_p \ge x/(\log x)^{\varpi}$, their (b.1), and a prime number theorem for products of primes weighted by $b_n$, their (b.2) \cite[\S4.2]{FM}. For given $\varpi > 1$ and $B$, the pairs $((a_n), (b_n))$ with all these properties form their class $\Psi$ \cite[Definition~4.7]{FM}. $C^-(\gamma, \theta, \nu)$ is the supremum of the constants $C$ such that for every $\varpi > 1$ there is $B$ with $\sum_p a_p \ge C \sum_p b_p$ for every pair in $\Psi$ and every sufficiently large $x$, and $C^+(\gamma, \theta, \nu)$ is defined symmetrically \cite[Definition~4.8]{FM}. For $0 \le \alpha < 1$ we add \emph{Hooley's condition at level $\alpha$}. Hooley suggested postulating ``that the parity bias in arithmetic progressions is small on average'', and Ford records it in the form $\sum_{d\le x^\alpha} |\sum_{n\le x,\, d\mid n} \mu(n)a_n| \ll_B x(\log x)^{-B}$ \cite[p.~1665]{Ford05}. Since FM's comparison sequence $b_n$ is general, we impose the condition on $w = a - b$: \[ (\Hoo_\alpha)\qquad \sum_{d \le x^\alpha} \Bigl| \sum_{\substack{x/2 < n \le x\\ d \mid n}} \mu(n)\, w_n \Bigr| \le \frac{x}{\log^B x}. \] The term $d = 1$ is the global parity condition \[ (\mathrm{Par})\qquad \Bigl|\sum_{x/2 < n \le x} \mu(n)\, w_n\Bigr| \le \frac{x}{\log^B x}, \] so $(\Hoo_0)$ is $(\mathrm{Par})$. We write $C^\pm(\gamma, \theta, \nu \mid \Hoo_\alpha)$ for the constants of \cite[Definitions~4.7 and~4.8]{FM} with $(\Hoo_\alpha)$ adjoined to (I) and (II), with FM's order of quantifiers ($B$ large, then $x$ large). Adjoining a condition can only shrink the class of sequences, so $C^-(\gamma, \theta, \nu \mid \Hoo_\alpha) \ge C^-(\gamma, \theta, \nu \mid \mathrm{Par}) \ge C^-(\gamma, \theta, \nu)$ for every $0 \le \alpha < 1$. All our results have $\gamma=1/2$ and $\theta=0$. Our prime-free constructions use $b_n=1$ on $(x/2,x]$: we give a family $a_n=a_n^{(x)}\ge0$, bounded uniformly in $x$ and independent of $B$, that for every $B>0$ and every sufficiently large $x$ satisfies (I), (II) on $1<\ell\le x^\nu$, and $(\Hoo_\alpha)$, with $a_p=0$ at every prime in $(x/2,x]$. Condition (w) holds automatically for these bounded families, and (b.1) and (b.2) hold for $b_n=1$ by \cite[Lemma~4.6]{FM}, with $q=1$, $y=x/2$. Thus $C^-(1/2,0,\nu\mid\Hoo_\alpha)=0$ for the parameters of these constructions. Theorem~\ref{thm:B} instead applies to general comparison sequences satisfying (b.1) and (b.2), and for $\nu < 1/3$ uses a sieve argument with an additional parity term. With $b_n = 1$, Hooley's form for $a_n$ and $(\Hoo_\alpha)$ for $w_n$ are equivalent for every fixed $\alpha < 1$, because the baseline sums are small. Write $n = dm$. The inner sum for $b$ is $\mu(d)$ times a difference of two sums $\sum_{m \le y,\,(m,d)=1}\mu(m) = \sum_{k \le y,\ k \mid d^\infty} M(y/k)$, with $M$ Mertens' function and $y = x/d$, $x/(2d)$. Split at $k = \sqrt y$; use $|M(y)| \ll y\,e^{-c\sqrt{\log y}}$ below the split and Rankin's trick above it. The total over $d \le x^\alpha$ is $\ll x \log x \log\log x\; e^{-c\sqrt{\log x}} + x^{3/4 + \alpha/4 + o(1)}$. \subsection{The continuum model} Let $0 < \eta < 1/2$. A squarefree integer $n$ is \emph{$\eta$-rough} if every prime factor of $n$ is at least $n^\eta$. For a modulus $d \le x^\alpha$ we use the scale $x$ instead: $d$ is $\eta$-rough if every prime factor of $d$ is at least $(x/2)^\eta$. Such an $n$ with prime factors $p_1 < \dots < p_k$ has the \emph{vector} $v(n) = (\log p_1/\log n, \dots, \log p_k/\log n)$, which lies on \[ \Delta_k(\eta) = \{u \in \mathbb R^k : \eta \le u_1 \le \dots \le u_k,\ u_1 + \dots + u_k = 1\}. \] A \emph{continuum function} is a family $f = (f_k)_{k \ge 1}$ of bounded functions on these simplices; we write $f(u)$ for the value at a vector $u$ of any length, and $f(1)$ for its value at the vector $(1)$ of a prime. The perturbations we construct have the form $w_n = F(v(n))$ on $\eta$-rough squarefree $n$, and $w_n = 0$ otherwise (\S\ref{sec:lift}). FM translate Type I information at level $\gamma$ into integral relations, their (6.2), and show that these are equivalent to a \emph{fragmentation relation}, their (6.3) \cite[Theorem~6.4]{FM}: the values of $f$ on \emph{free} vectors, those with every component in $[\eta, 1 - \gamma)$, may be prescribed arbitrarily, and (6.3) then determines $f$ on every other vector. We always take $\gamma = 1/2$. A vector that is not free has exactly one component $\ge 1/2$, except the vector $E = (1/2, 1/2)$, which has two. For a vector $(a, \beta)$ with one large component $a \ge 1/2$, (6.3) expresses $f(a, \beta)$ through the free values on all fragmentations of $a$ into components below $1/2$, weighted by FM's truncated Linnik kernel $L_{1/2}$: \[ f(\beta, a) = a \sum_{k \ge 2} \int_{\substack{u_1 < \dots < u_k,\ u_1 \ge \eta\\ u_1 + \dots + u_k = a}} L_{1/2}(u)\, f(\beta, u)\, \frac{du}{u_1 \cdots u_k}, \] with FM's projection measure. In the twisted functionals below, the factor $a$ cancels the weight $1/a$ of the large component, so the kernels of Appendix~\ref{app:kernels} carry no factor $a$; values at vectors with a large component do. We use the following closed form of $L_{1/2}$. \begin{lemma}[The truncated Linnik kernel]\label{lem:K} Let $k \ge 1$, let $0 < u_i < 1/2$ for $1 \le i \le k$, let $a = u_1 + \dots + u_k \le 1$, and let $N_{1/2}(u)$ be the number of unordered partitions of $\{1, \dots, k\}$ into two blocks each of $u$-sum $1/2$. Then \[ L_{1/2}(u) = \mathbf 1_{k = 1} + (-1)^{k+1} \sum_{\substack{B \subseteq \{1, \dots, k\}\\ u_B \ge 1/2}} (-1)^{|B|} - \mathbf 1_{a = 1}\, N_{1/2}(u), \qquad u_B := \sum_{i \in B} u_i . \] \end{lemma} This is the case $c = 1/2$ of the generating-function identity (5.3) of \cite[\S5]{FM}: expand $\log V(z)$ with $V(z) = \prod_i (1 + z_i) - \sum_{u_A \ge 1/2} z^A$ and compare the coefficients of $z_1 \cdots z_k$. We give the direct count. \begin{proof} An unordered partition into $r$ blocks has weight $(-1)^{r-1}(r-1)!$ in FM's (5.1), and the weights of all partitions sum to $\mathbf 1_{k = 1}$. For a prescribed block $B$ with $u_B \ge 1/2$, summing over the partitions of its complement gives $(-1)^{k - |B|}$ (a partition containing $B$ with $r'$ further blocks has weight $(-1)^{r'} r'!$, and $\sum_{r'} (-1)^{r'} r'!\, S(n, r') = (-1)^n$ for the Stirling numbers $S(n, r')$, since $\sum_r (-1)^r (e^z - 1)^r = e^{-z}$); subtract these contributions. Two blocks of sum at least $1/2$ occur only when $a = 1$, where each has sum exactly $1/2$; inclusion and exclusion adds back their weight $-1$ once for each unordered pair. Three such blocks are impossible. \end{proof} \subsection{The twisted functionals} For a continuum function $F$ and a tuple $s = (s_1, \dots, s_t)$ with $t \ge 1$, $s_i \ge \eta$ and $|s| := s_1 + \dots + s_t < 1$, put \begin{align*} T_t(F)(s) &:= \sum_{k > t} (-1)^k \int_{\substack{z_{t+1} < \dots < z_k\\ z_{t+1} + \dots + z_k = 1 - |s|}} F(s, z)\, \frac{dz}{z_{t+1} \cdots z_k},\\ T_0(F) &:= \sum_{k \ge 1} (-1)^k \int_{\substack{z_1 < \dots < z_k\\ z_1 + \dots + z_k = 1}} F(z)\, \frac{dz}{z_1 \cdots z_k}, \end{align*} with FM's projection measure on the hyperplanes, and $F(s, z)$ the value at the vector with the components of $s$ and $z$. The term $k = 1$ of $T_0$ is $-F(1)$. These are the continuum forms of the twisted sums in $(\Hoo_\alpha)$. For $d = p_1 \cdots p_t$ with every $p_i \ge (x/2)^\eta$, put $\lambda_i = \log p_i/\log x$ and $\lambda = \lambda_1 + \dots + \lambda_t$. The main term of $\sum_{d \mid n} \mu(n) w_n$ is then \[ \int_{1 - \lambda + \log(1/2)/\log x}^{1 - \lambda} x^w\, q^{-1}\, T_t(F)\bigl(\lambda_1/q, \dots, \lambda_t/q\bigr)\, dw, \qquad q = \lambda + w \] (Lemma~\ref{lem:L3} below), and $T_0$ is the case $d = 1$. The \emph{untwisted} functionals are the same sums without the sign $(-1)^k$. For every $F$ satisfying (6.2) at level $1/2$, the untwisted $T_0$ vanishes, and the untwisted $T_t(F)(s)$ vanishes for $t\ge1$ and $|s|\le1/2$. This is the Type I condition, and the certificates check it as a control, to the extent stated in \S\ref{sec:certificates}. For $\eta$-rough moduli up to $x^\alpha$, the tuples that occur have $|s| = \lambda/q \le \alpha/(1 - \log 2/\log x)$, since $q \ge 1 - \log 2/\log x$, and tuples with some $s_i < \eta$ contribute nothing. So it suffices that $T_0(F) = 0$ and $T_t(F)(s) = 0$ for every $t \ge 1$ and every tuple with $s_i \ge \eta$ and $|s| \le \alpha''$, for some fixed $\alpha'' > \alpha$; the constructions of \S\ref{sec:construction} give this with $\alpha'' = \alpha'$. A tuple of $t$ components $\ge \eta$ has $|s| \ge t\eta$, so only $t \le \alpha''/\eta$ occurs.