\section{Introduction}\label{sec:intro} \subsection{Type I, Type II and parity} Let $(a_n)$ be a nonnegative sequence supported on $(x/2, x]$, compared with a sequence $(b_n)$ whose prime sum is understood. A sieve method bounds $\sum_p a_p$ in terms of $\sum_p b_p$, using only how well $w = a - b$ is distributed: in arithmetic progressions (Type I information) and against bilinear forms (Type II information). Ford and Maynard \cite{FM} recently set this up as a general framework. Type I information (I) holds up to level $x^\gamma$, and Type II information (II) holds for one factor in $((x/2)^\theta, x^{\theta + \nu}]$. They define the best constants $C^-(\gamma, \theta, \nu) \le C^+(\gamma, \theta, \nu)$ with $(C^- + o(1)) \sum_p b_p \le \sum_p a_p \le (C^+ + o(1)) \sum_p b_p$ \cite[Definition~4.8]{FM}; \S\ref{sec:setting} recalls the definitions. Type II information is indispensable. With $\nu = 0$, Selberg's sequence $a_n = 1 + \lambda(n)$, with $b_n = 1$ and $\lambda$ the Liouville function, satisfies (I) for every $\gamma < 1$ and has no primes \cite{Selberg}, \cite[\S1]{FM}. This is the parity phenomenon. Ford and Maynard show that a small Type II range does not suffice either: for every $\gamma < 1$ there is $\nu_0(\gamma) > 0$ such that $C^-(\gamma, \theta, \nu) = 0$ for $\nu \le \nu_0(\gamma)$, which they describe as ``the first examples (at least with $\gamma \geqslant 1/2$ and $\nu > 0$)'' of prime-free sequences with nontrivial Type I and Type II information \cite[Theorem~2.1 and \S2.1]{FM}. In the classical case $\gamma = 1/2$, $\theta = 0$, Ford and Maynard state that primes are forced at $\nu = 0.16623$ and for every $\nu \ge 1/5$, and show $C^-(1/2, 0, 0.1616) = 0$ \cite[Theorems~2.2 and 2.7, the proofs of 2.7(b), (c), and the case $\nu \ge 1 - \gamma$ of \S10.1]{FM} (for $\nu \ge 1/3$ see also \S\ref{ssec:large-nu}). Ford has since reported a prime-free example at $\nu = 0.163$ \cite[p.~2728]{Ford25}. \subsection{Hooley's condition} Selberg's example is extremely biased in its parity: $\sum_n \mu(n) a_n$ is of the order of $x$. Under the density, growth and squarefree-support hypotheses of their Theorem~1, Friedlander and Iwaniec combine Type~I information at a level $x^{2/3} < D < x$ with their M\"obius bilinear condition to obtain the prime-sum formula with relative error $O(\log\delta/\log\Delta)$. It is an asymptotic when their bilinear-range parameters satisfy $\log\delta = o(\log\Delta)$ \cite[Theorem~1]{FI98}. Ford asked whether, for each Type I level $\vartheta < 1$, there are sequences with Type I information at level $\vartheta$ and no global parity bias, $\sum_{n \le x} \mu(n) a_n \ll_B x(\log x)^{-B}$, for which Bombieri's asymptotic formulas $\sum_{n \le x} a_n \Lambda_k(n) \sim k x (\log x)^{k-1}$, with $\Lambda_k = \mu \ast \log^k$, fail. He constructed such sequences for which they fail for every $k \ge 2$ and hold for $k = 1$, the case of primes \cite[p.~1665 and Theorem~2]{Ford05}. He called it an interesting problem to replace global parity balance by a stronger condition, but one weaker than Friedlander and Iwaniec's, and as one possibility recorded Hooley's suggestion ``that the parity bias in arithmetic progressions is small on average'', in the form \[ \sum_{d \le x^\alpha} \Bigl| \sum_{\substack{n \le x\\ d \mid n}} \mu(n) a_n \Bigr| \ll_B \frac{x}{(\log x)^B} \] \cite[p.~1665]{Ford05}. We write $(\Hoo_\alpha)$ for this condition, imposed on $w = a - b$ over $(x/2, x]$, and $C^-(\gamma, \theta, \nu \mid \Hoo_\alpha)$ for Ford and Maynard's constant over the sequences that also satisfy it (\S\ref{sec:setting}). The term $d = 1$ alone is the global parity condition $(\mathrm{Par})$, a dyadic form of Ford's condition. Ford's question involves no Type II information; this paper studies what $(\mathrm{Par})$ and $(\Hoo_\alpha)$ do when Type II information is added. Ford and Maynard also ask for ``a theory which can incorporate such additional arithmetic information'' (their examples are trilinear estimates, density assumptions and estimates for special coefficients), and add that they ``have not attempted such a generalization'' \cite[\S4.6]{FM}. Twisted information of related kinds is known to break the parity barrier. For Goldbach's problem, Huang and Li fix $A > 0$ and $0 < \theta < 1$, and assume the Elliott--Halberstam conjecture for the primes at level $N^\theta(\log N)^{2A+8}$ together with its variant twisted by the Möbius function at level $N^{1-\theta}$: for every $B > 0$, \[ \sum_{q \le N^{1-\theta}}\, \max_{y < N}\, \max_{(a, q) = 1} \biggl| \sum_{\substack{n \le y\\ n \equiv a \bmod q}} \Lambda(n)\mu(N - n) - \frac{1}{\varphi(q)} \sum_{n \le y} \Lambda(n)\mu(N - n) \biggr| \ll_B \frac{N}{(\log N)^B}. \] Under these hypotheses $\sum_{n < N} \Lambda(n)\Lambda(N - n) \ge \mathfrak S(N)(1 - \mathcal A(N))N + O(N(\log N)^{-A})$ for every even $N$, where $\mathfrak S(N)$ is the singular series and $\mathcal A(N) = \prod_{p \nmid N,\, p > 2}\bigl(1 - \frac{1}{p(p-1)}\bigr)$ \cite[Theorem~1]{HL}. They deduce Goldbach's conjecture for all large even $N$ from the same hypotheses with the first at level $N^\theta(\log N)^{12+\epsilon}$, and, since the Bombieri--Vinogradov theorem supplies that hypothesis for every $\theta < 1/2$, from the twisted hypothesis at some level above $1/2$ alone \cite[Corollary~1]{HL}. Moya \cite{Moya} reports that the divisor switch in \cite[(18)]{HL} drops a moving endpoint, $n < N - \alpha k$ with $\alpha$ their divisor cut-off, and states that the conclusion holds under the original hypotheses with the endpoint restored; this is one reason the interval maximum matters. For a fixed nonzero even integer $h$, Murty and Vatwani assume $EH_\Lambda(x^\theta(\log x)^C)$ and $EH_{\mu_h}(x^{1-\theta})$ for a fixed $\theta < 1$ and a sufficiently large fixed $C$, where the second hypothesis is their centred progression estimate for $\Lambda(n)\mu(n + h)$. They prove $\sum_{n \le x} \Lambda(n)\Lambda(n + h) \ge (1 - o(1))\mathfrak S(h)(1 - \mathcal A(h))x$, with $\mathcal A(h) = \prod_{p \nmid h,\ p > 2}(1 - 1/(p(p-1)))$ \cite[Theorem~1.1]{MV}; only positive arguments of the arithmetic functions are included. These hypotheses concern particular sequences. They subtract the full initial-interval sum divided by $\varphi(q)$ and take maxima over reduced residue classes and initial intervals. They are not the condition $(\Hoo_\alpha)$ for a general sequence. For general sequences, Iwaniec's sieve for primes and Friedlander and Iwaniec's inequality combine Type I information at level $x^{1-\delta}$ with a bilinear form in which the Möbius function weights a factor in $(x^\delta, x^{2\delta}]$ \cite[\S6, Proposition, p.~296]{Iw06}, \cite[Proposition~6.1]{FI22} (a variant of \cite[Theorem~18.5]{FIOpera}); Vatwani treats twin primes under an Elliott--Halberstam-type conjecture for the primes $p$ with $p + 2$ squarefree, at every level below $1$, together with a Chowla--Elliott-type bound $\sum_{d < x^\theta} |\sum_{p \le x,\ d \mid p + 2} \mu(p + 2)| \ll x(\log x)^{-2}$ for every $\theta < 1/2$, which has the shape of $(\Hoo_\theta)$ for that particular sequence \cite[Chapter~7]{Vatwani}. Whether $(\Hoo_\alpha)$ at a level $\alpha > 1/2$ forces primes for general sequences in Ford and Maynard's framework is not addressed in this paper beyond Theorem~\ref{thm:B}, by which, with $\gamma = 1/2$ and $\theta = 0$, the term $d = 1$ of $(\Hoo_\alpha)$ alone already forces primes once $\nu \ge 0.116$. In the plane of the Type II range $\nu$ and the twisted level $\alpha$, with Type I information at level $1/2$, Ford and Maynard state results that force primes at every $\alpha$ for $\nu = 0.16623$ and for every $\nu \ge 1/5$ \cite[Theorems~2.2 and 2.7(a), the proof of 2.7(b), and the case $\nu \ge 1 - \gamma$ of \S10.1]{FM} (for $\nu \ge 1/3$ see also \S\ref{ssec:large-nu}). This paper asks how far into the rest of this plane prime-free sequences reach. \subsection{Results} Our first theorem gives prime-free sequences with Type I information at level $1/2$, a positive Type II range, and Hooley's condition at every twisted level below $4/15$. \begin{theoremA}[The ladder]\label{thm:A} Let $(\eta, \alpha')$ be one of \[ (1/11,\ 1/10), \quad (1/14,\ 1/8), \quad (1/18,\ 1/6), \quad (1/20,\ 1/5), \quad (1/30,\ 4/15). \] Fix $0 < \nu < \eta$ and $0 \le \alpha < \alpha'$. There is a uniformly bounded family of sequences $a_n = a^{(x)}_n \ge 0$ supported on $(x/2, x]$ with the following property. For every $B > 0$ and every sufficiently large $x$, the sequence $w_n = a_n - 1$ satisfies (I) with $\gamma = 1/2$, (II) with $\theta = 0$ on $1 < \ell \le x^\nu$, and $(\Hoo_\alpha)$. Moreover, $a_p = 0$ for every prime $p \in (x/2, x]$. In particular, $C^-(1/2, 0, \nu \mid \Hoo_\alpha) = 0$. \end{theoremA} The five cases are the corners of a staircase in the $(\nu, \alpha)$-plane. In our constructions a larger twisted level came at the cost of a shorter Type II range; we do not know whether this trade-off is forced. The case $(1/30, 1/4)$, contained in $(1/30, 4/15)$, is proved on the way; it is the input of one of the Goldbach consequences in \S\ref{sec:goldbach}. The simplest construction, $(\eta, \alpha') = (1/11, 1/11)$, is implied by the case $(1/11, 1/10)$ and is described in Remark~\ref{rem:base} for illustration. Our second theorem says that the term $d = 1$ alone, global parity balance, already forces primes once $\nu \ge 0.116$. Without it, Ford and Maynard's examples give $C^-(1/2, 0, \nu) = 0$ for every $\nu \le 0.1616$ \cite[Theorem~2.7(c)]{FM}: their construction at $\nu = 0.16169$ satisfies (I) and (II) there, hence at every smaller $\nu$, where the Type~II range only shrinks, and its $b_n = 1$ and bounded $a_n$ give the remaining conditions \cite[Theorem~6.3(b) and Lemma~4.6]{FM}. So on $[0.116, 0.1616]$ the parity condition changes the answer. \begin{theoremA}[Parity forces primes from $\nu = 0.116$]\label{thm:B} \begin{enumerate} \item[(a)] For $0.116 \le \nu < 1/3$, \[ C^-(1/2, 0, \nu \mid \mathrm{Par}) \;\ge\; 1 - 2I_4(\nu) - 4I_5(\nu) - \tfrac{20}{3} I_6(\nu) - \tfrac{25}{3} I_7(\nu) - \tfrac{35}{3} I_8(\nu) \;>\; 0, \] where $I_k(\nu) = \frac{1}{k!}\int \frac{dx_1 \cdots dx_{k-1}}{x_1 \cdots x_k}$ over $x_i \in (\nu, 1 - \nu)$ with $x_1 + \dots + x_k = 1$, and $I_k = 0$ if $k\nu \ge 1$; in particular $I_8 = 0$ for $\nu \ge 1/8$. The bound is at least $0.0350$ throughout the range, at least $0.2892$ for $\nu \ge 1/8$, at least $0.6198$ for $\nu \ge 1/7$, and at least $0.7072$ for $\nu \ge 3/20$. \item[(b)] For $1/4<\nu<1/3$, (w), (I), (II) and $(\mathrm{Par})$ give $\sum_{x/2
0$ for every $0 \le \alpha < 1$ and every $0.116 \le \nu \le 1$: for $\nu < 1/3$ this follows from (a), since the term $d = 1$ of $(\Hoo_\alpha)$ is $(\mathrm{Par})$; for $1/3 \le \nu \le 1$, (w), (I) and (II) with $\gamma = 1/2$ and $\theta = 0$ already give $\sum_{x/2
(x/2)^\eta$. Ford and Maynard's examples for their Theorem~2.1 are built from their modified Liouville functions, products over the components of a one-variable function \cite[\S\S9.1--9.2]{FM}. The main term is $(-1)^k$ on vectors whose $k$ components all lie below $1 - \gamma$; a correction term also depends on the sizes of the components. These examples are built for Type~I and Type~II information and have no reason to satisfy twisted conditions. We add, for each number $j$ of ``small'' components (those in $[\eta, \alpha']$), a function $h_j$ of their total size. The twisted rows at moduli $p_1 \cdots p_t$ are then linear in $h_t$ with a nonvanishing diagonal coefficient, so they can be solved exactly, top-down in $t$ (\S\ref{sec:construction}). A linear program chooses the remaining constants, and a rigorous enclosure certifies, with a margin, the inequalities that make $a_n \ge 0$ once $f$ is normalised so that $a_p = 0$. Ford and Maynard's construction procedure (the proof of \cite[Theorem~6.3]{FM}), with their mask, which makes the function vanish on every vector with a subsum near $1/2$ or just above $\nu$ (the vector $E = (1/2, 1/2)$, where the certificate asserts no bound, among them), followed by a re-solve of the forced functions and a normalisation, then turns the continuum function into sequences with $a_p = 0$ exactly (\S\ref{sec:lift}). \subsection{Limits} The construction meets an obstruction past $1/3$. For $\alpha > 1/3$, some twisted rows involve moduli that no vector of the support can complete with rough components alone, and we prove that the forced functions $h_j$ do not enter these rows (Lemma~\ref{rem:dead}). Floating-point experiments (for the profile $w_k = 1$ at $\eta = 1/36$, on a grid of step $1/72$) suggest that the method cannot pass $1/3$, but that part is a conjecture about this family of constructions (Conjecture~\ref{conj:wall}), not a statement about Hooley's condition: on three of the four grids we could compute, a discretised version of the problem with fully free data, capped at $4$ times the baseline and with Type~I information that includes moduli of size exactly $x^{1/2}$, does get past $1/3$ (with exact certificates on two of them). How close to $1/3$ the family can get is not known: the certified cases reach $4/15$, and in floating point, on one grid, the family's values just past $1/3$ stay far above $-1$ for every box and bin choice we tried, while for $\eta \le 1/36$ richer data that also see the largest small component come within $0.05$ of $-1$ at the least level computed at or past $1/3$ on each grid (below it on two of the four grids under both boundary conventions tested, with the other two unresolved under the tested conventions and boxes), from which we draw no conclusion. \subsection{Consequences for Goldbach} The constructions reflect to Goldbach's problem (\S\ref{sec:goldbach}). For each case of Table~\ref{tab:goldbach} and every large even $N$ there is a nonnegative sequence on $[1, N)$ that vanishes on every prime $m \ge N/(\log N)^C$, and that agrees with a presieved local model of $\Lambda(N - m)$ in five respects: Type~I sums at the Bombieri--Vinogradov level, small-factor Type~II sums (exactly), Hooley's twisted condition at level $\alpha$, parity sums over every interval, and roughness comparisons up to $(N/(2(\log N)^C))^\eta$. So these inputs do not by themselves force a Goldbach-scale count, even with Hooley's condition granted up to any fixed level $\alpha < 4/15$ (with Type~II below $N^{1/30}$). The Type~II identity, the twisted condition and the parity sums at the precision $N(\log N)^{-A}$ are properties of the reflected sequence. In the arbitrary-coefficient form the reflection uses, the true sequence fails the first, as it fails the exact roughness identity, and it is not known to have the other two. For it, even $\sum_{p < N} \mu(N - p) = o(\pi(N))$ has been obtained only recently and conditionally \cite{Guo}, building on L\"u, Wang and Wu's work on $\sum_{p \le x} \mu(p + 2)$ \cite{LWW}. \S\ref{sec:goldbach} says what the audits of the reflection checked, and each case carries the certification status of its construction (\S\ref{ssec:status}). \subsection{Prior work and credit} The continuum model, of vectors of prime factors and Type I-compatible functions \cite[Definitions~5.1 and~6.2]{FM}, the fragmentation relation \cite[(6.3), Theorem~6.4]{FM}, the mask and the construction theorem \cite[Theorem~6.3 and its proof in \S6.2]{FM} are Ford and Maynard's. Ford's constructions for Bombieri's asymptotic sieve \cite{Bombieri76} rest on a related observation: his perturbations $a_n - 1$ are supported on integers whose prime factors lie in windows near $x^{i/M}$, and moduli outside these windows see only the baseline \cite[(3.2) and p.~1669; p.~1673]{Ford05}. Ford remarks that the sequences constructed for his Theorem~1 satisfy Hooley's condition for every fixed $0 \le \alpha < 1-\vartheta$, where $\vartheta$ (his $\nu$) is the Type~I level \cite[pp.~1663, 1665--1666]{Ford05}. This does not hold as printed, and we do not use it: for those sequences, $\bigl|\sum_{n\le x}\mu(n)a_n\bigr| \gg x/\log x$ for arbitrarily large $x$, so the $d=1$ term already violates the condition \cite[p.~1667; (3.9)--(3.10), pp.~1671--1672; p.~1673]{Ford05}; Ford notes at his (1.5) that these sequences ``all exhibit a `global parity bias'\,'' \cite[p.~1665]{Ford05}. The lower bound in Theorem~\ref{thm:B}(a) follows the proof of Ford and Maynard's sieve bound \cite[Theorem~7.3]{FM}, with one additional parity term and a separate sign condition at non-squarefree integers. We are not aware of an earlier prime-free construction satisfying the twisted condition, including its $d = 1$ term. \subsection{Organisation} \S\ref{sec:setting} sets up the framework and the continuum model. \S\ref{sec:construction} gives the construction and \S\ref{sec:lift} the passage to sequences. \S\ref{sec:certificates} describes the certificates and their audits, and \S\ref{sec:rungs} the individual cases. \S\ref{sec:converse} proves Theorem~\ref{thm:B}. \S\ref{sec:limits} discusses the obstruction past $1/3$ and Conjecture~\ref{conj:wall}, and \S\ref{sec:goldbach} the consequences for Goldbach. \S\ref{sec:computations} and the supplement describe how to reproduce the computations.