\section{Limits of the construction}\label{sec:limits} The certified cases stop at $4/15$ for a computational reason. Past $1/3$ the family \eqref{eq:free} meets a structural obstruction, and we conjecture that it cannot pass $1/3$; how close to $1/3$ it can get is not known. In floating-point experiments the obstruction appears to belong to the family rather than to the problem: on three of the four grids we could compute, a discretised version of the problem with fully free data (Type~I rows up to and including level $1/2$, values at most $4$ times the baseline; see below) gets past $1/3$, while for richer data that see the largest small component and $\eta \le 1/36$ the values at the least level computed at or past $1/3$ on each grid lie within $0.05$ of $-1$ (below it on two of the four grids under both boundary conventions tested, with the other two unresolved under the tested conventions and boxes), and at the next grid point between $-0.96$ and $-0.78$. We keep the proved, measured and conjectural parts apart. \subsection*{The next case} An uncertified numerical candidate occurs at the natural next case, $(\eta, \alpha') = (1/36, 17/60)$, which has $m^* = 10$ forced levels: with the profile $w_2 = 1$, $w_3 = -2$, constants designed on a grid of step $1/1800$ and checked on grids of step $1/1440$ and $1/2160$, after re-forcing the functions and adjusting two class constants to impose $T_0 = 0$, give a positive numerical margin $1 - \min f / f(1)$ of a few hundredths (between $0.03$ and $0.05$ at the two resolutions; floating point, from one design code, not independently recomputed; the supplement holds the constants but not the design code or its evaluation logs). We do not claim this case. With the second-order engine of \S\ref{sec:certificates} as it stands, a certificate is out of reach at affordable cost, for two reasons. First, affected one-step couplings (the terms $C_{j,1}$ of \eqref{eq:row}) have a first-order enclosure loss when a threshold window straddles the jump of the rough kernel at $\alpha'$; at fixed block width this loss is not removed by increasing $M$, and increasing the number of blocks reduces it in the measured runs. A control run at $(1/30, 4/15)$ shows no comparable loss. The likely reason, which we have not proved, is that such cells occur only when $2\alpha' - 1/2 > \eta$, while at $(1/30, 4/15)$, $2\alpha' - 1/2 = \eta$ exactly. Second, the widths produced by the second-order engine compound down the forcing chain, by a measured factor of about $3$ per level from level $6$ to level $1$: at $M = 3600$ the width of $h_1$ is about $3.6$, against a margin of a few hundredths. Between $M = 3600$ and $5400$ the widths at levels $3$ to $5$ shrank approximately as $M^{-1.5}$; assuming a quadratic rate would be optimistic, and we do not supply a cost projection. The measurements suggest treating the jump exactly and reducing the propagated enclosure widths; they do not establish that these changes are necessary or sufficient. The certified cases currently end at $4/15$. \subsection*{The obstruction past $1/3$} The construction of \S\ref{sec:construction} adds to Ford and Maynard's class constants one forced function $h_j$ for each number $j$ of small components, a function of their total size. We conjecture that such constructions cannot pass twisted level $1/3$ (Conjecture~\ref{conj:wall} below). The proved part is the following. \begin{lemma}[Rows that the forced functions cannot reach]\label{rem:dead} Let $1/3 \le \alpha' < 1/2$, and let $s$ be a tuple of small components with total $\sigma \in [1 - 2\alpha', \alpha']$, an interval that is nonempty exactly when $\alpha' \ge 1/3$. \begin{enumerate} \item[(a)] Every free vector that contains the components of $s$ has at most one rough component, and its small total exceeds $1/2$. \item[(b)] Let $g$ be a bounded function on free vectors that vanishes on those with small total greater than $1/2$. Then the extension of $g$ by (6.3) vanishes at every vector that contains the components of $s$, and so $T_t(g)(s) = 0$. \end{enumerate} In particular, for free data \eqref{eq:free} the functions $h_j$ do not enter the twisted rows $T_t(f)(s)$ at these tuples: only the class constants do. \end{lemma} \begin{proof} (a) Let $u$ be a free vector that contains the components of $s$, and let $z$ consist of its other components. They total $1 - \sigma \in [1 - \alpha', 2\alpha']$, and each lies in $[\eta, 1/2)$, so each is small or rough. Two rough components of $z$ would total more than $2\alpha' \ge 1 - \sigma$, so $z$ has at most one. If $z$ had no small component, it would be a single rough component of size $1 - \sigma \ge 1 - \alpha' > 1/2$, which is impossible. So the small total of $u$ is $1$ if $z$ has no rough component, and $1 - r > 1/2$ if it has one, $r$. (b) By (a), $g$ vanishes at every free vector that contains the components of $s$. A vector that contains them and is not free is not $E$, whose components are $1/2 > \alpha'$, so it has exactly one large component $a$. Relation (6.3) expresses its value through the free values at the vectors obtained by fragmenting $a$ into components below $1/2$. Each of these is a free vector that still contains the components of $s$, so the extension of $g$ vanishes there as well. The functional $T_t(\cdot)(s)$ integrates the values at the vectors $(s, z)$, all of which contain the components of $s$, so $T_t(g)(s) = 0$. For the last statement, apply (b) to the forced part $\mathbf 1[1 \le j \le m^*,\ \tau \le \alpha']\, w_k h_j(\tau)$ of \eqref{eq:free}, which vanishes when $\tau > \alpha'$, and use the linearity of the extension and of $T_t$. \end{proof} The class constants enter these rows through smoothing kernels, so for the free data \eqref{eq:free} the rows of Lemma~\ref{rem:dead} are linear conditions on the finitely many constants alone, one for each $\sigma$ in an interval. (In the grid programmes below, the rows at $\sigma = 1 - 2\alpha'$ also contain the added function of their own label, through the closure node with two rough components equal to $\alpha'$; see the README.) The rest of this section reports floating-point experiments. Except for the rigorous certificates named below, every value in it is a floating-point solution of a finite linear programme: evidence, not proof. In the experiments on families, the programme minimises $f(1)$ over the class constants and the added functions, tabulated on a grid of step $1/N$, subject to the twisted rows and to $f \ge -1$ at every generated non-prime grid point, including the closure-labelled free nodes and the extended points described in the README, except at $E$ read as a rough component plus a large one (there is no upper positivity condition). The problem itself restricts the data only through $f \ge -1$ and boundedness, so each programme carries an explicit box, which we state and vary: the class constants lie in $[-1, U]$ and the added functions are bounded by $C$ in absolute value. For $\alpha' \ge 1/3$, allowing the constants of the classes whose free vectors all carry added data to go down to $-L < -1$ adds nothing to the family itself. Adding a constant $\delta_j$ to $c[j, 2, b]$ for every bin $b$ and subtracting $\delta_j$ from the added function of level $j$ leaves $f$ unchanged at every free vector, and the same holds for the classes with one rough component. So that box gives no more than the box with $U + L - 1$ and $C + L - 1$ in place of $U$ and $C$. On the grid this fails at least at the nodes where the one rough component equals $1/2$ and at the $G_j$ columns that the README lists as omitted, which carry quadrature weight but no added variable; for $c[j, 2, b]$ the exchange is exact on the grids we checked ($\alpha' = 24/72$ and $25/72$). We therefore bound every constant below by $-1$. Unless stated otherwise, the family programmes below use $U = 4$, with $C = 50$ for the richer families. All the families below have free data of the form \begin{equation}\label{eq:freeM} f(u) = c[j, k, b] + \mathbf 1[\tau \le \alpha']\, H_j(\tau, \mathfrak m) + \mathbf 1[k = 1]\, G_j(\tau, \mathfrak m), \end{equation} with $j$, $k$, $b$ and $\tau$ as in \eqref{eq:free} and $\mathfrak m$ the largest small component: the forced family \eqref{eq:free} with $w_k = 1$ is the case where $H_j$ does not depend on $\mathfrak m$ and $G_j = 0$, and the family with $G_j$ the case where neither depends on $\mathfrak m$. The values are tabulated at the grid points $i/N$, and every integral in the rows and in (6.3) is a composite trapezoid sum over the grid points of its closed range, so that a component equal to $\alpha'$ counts half as small and half as rough, and a component equal to $1/2$ half as rough and half as large. The kernels of (6.3) are evaluated at the endpoint $a = 1/2$ of the large component by their limits from above, for the class constants and the added functions alike, so that each programme depends on the data only through the values of $f$ at the grid points. The remaining grid conventions of these programmes (closure points, the vector $E$, and the added functions they leave out) are listed in the supplement's experiments README, under ``Conventions of the family programmes''. Coefficients of absolute value below $10^{-9}$ are deleted from the matrices; on the resulting matrices the rows of Lemma~\ref{rem:dead} have right-hand-side tolerance $10^{-9}$, and the other twisted rows are equalities. Accepted solves have status Optimal, equality residuals, positivity violations and box violations at most $10^{-7}$, and residuals of the rows of Lemma~\ref{rem:dead} at most $10^{-9} + 10^{-10}$. A value below $-1$ is quoted as a numerical prime-free candidate only if it passes these checks; on the matrices before the deletion, the residuals of the retained twisted rows and the generated positivity violations are at most $5 \cdot 10^{-6}$ in every retained-row case quoted. In the explicitly stated comparison without the rows of Lemma~\ref{rem:dead}, those rows are exempt from the residual requirement. In every solution quoted, $f \ge -1 - 10^{-9}$ at every free grid point, the end points of the intervals included. Each value comes with a binary64 weak-duality lower bound for its programme, computed on the matrices before the deletion. A value above $-1$ whose bound is also above $-1$ is numerical evidence that the programme has no candidate. A separate check in an earlier discretisation of the forced family, on grids of step $1/360$ and $1/720$, agrees with the lemma: at the points checked, the coefficients of the forced functions in the rows of Lemma~\ref{rem:dead} are at most $4 \cdot 10^{-16}$, which is binary64 rounding, except within three grid steps of the endpoint $\sigma = 1 - 2\alpha'$, where the discretisation leaves errors of the order of the grid step. For the forced family \eqref{eq:free} with the profile $w_k = 1$ at $\eta = 1/36$, on a grid of step $1/72$, with two bins for the rough component (split at $30/72$ for $\alpha' \le 1/3$ and at $31/72$ past $1/3$) and the forced functions bounded by $10^4$, the best $f(1)$ is $-0.910$, $-0.778$, $-0.168$ and $-0.004$ at $\alpha' = 23/72$, $1/3$, $25/72$ and $13/36$ for $U = 4$; for $U = 1000$ each value changes by less than $2 \cdot 10^{-4}$. Without the rows of Lemma~\ref{rem:dead} the values at $25/72$ and $13/36$ are $-0.949$ and $-0.964$ ($U = 4$). On this grid, retaining those rows raises the objective sharply at the two tested points past $1/3$; this does not prove a bound for the continuum family. With bins of one grid step the values at $25/72$ and $13/36$ are $-0.258$ and $-0.186$ for $U = 4$, and $-0.265$ and $-0.187$ for $U = 1000$: the collapse persists, but its size depends on the bins. On a grid these values depend on the conventions at the boundaries, which are load-bearing there: the twisted rows include $\sigma = \alpha'$; Type~I enters only through (6.3), so there is no condition at a modulus of size exactly $x^{1/2}$; the positivity row at $E$ read as a rough component plus a large one is dropped; and a component equal to $\alpha'$ or $1/2$ is weighted half on each side. $E$ read as two rough components is the only vector of its class, whose constant enters only $T_0$, with coefficient $1/(2N)$. That constant is bounded below by $-1$ like every other, and the bound is active at every optimum reported in this section, so the parity row reads $T_0'x = 1/(2N)$, where $T_0'$ is $T_0$ without that term. Fixing the constant at $0$, so that $T_0'x = 0$, raises the values by $0.0029$ to $0.0033$ at the five points we tested ($25/72$ for the forced family, for the family with $G_j$ at two boxes and for the richer family, and $1/3$ for the richer family; see the README). These boundary conventions can materially affect the finite programmes. The runs reported here do not bound the change produced by another convention, or the error in locating a continuum obstruction. Under the stated conventions, the two tested forced-family values past $1/3$ remain far above $-1$. \begin{conjecture}\label{conj:wall} Let $0 < \eta < \alpha' < 1/2$, fix finitely many bins, and let $f$ be the extension by (6.3) of free data of the form \eqref{eq:free}, with any profile $(w_k)$, any $m^*$, any class constants and any bounded measurable functions $h_j$ on $[j\eta, \alpha']$. Then there is no $\alpha \in (1/3, \alpha']$ such that $f$ satisfies \eqref{eq:target} for some $m > 0$, $T_0(f) = 0$, and $T_t(f)(s) = 0$ for every $t \ge 1$ and every tuple $s$ with $s_i \ge \eta$ and $|s| \le \alpha$. \end{conjecture} Theorem~\ref{thm:lift} uses \eqref{eq:target} only at vectors with no subsum equal to $1/2$; the conjecture is stated for the full form, which is the form the experiments test. The conjecture rests on Lemma~\ref{rem:dead} and on the experiments above, which cover $\eta = 1/36$, one grid, $\alpha = \alpha'$ at two points, and the bins stated. The profile is no restriction there: for $\alpha' \ge 1/3$ a free vector with small total at most $\alpha'$ has exactly two rough components and small total below $1 - 2\alpha'$, so only $w_2$ and the values of $h_j$ on $[j\eta, 1 - 2\alpha')$ enter. The lemma applies to any bounded added data that vanish on free vectors with small total above $1/2$, not only to \eqref{eq:free}. It does not apply to data that vary with the small total on vectors with exactly one rough component $r$, where $\tau = 1 - r > 1/2$: completing a tuple of Lemma~\ref{rem:dead} by one rough component $r$ leaves small mass $1 - \sigma - r > 0$, so such data can enter those rows when the remaining mass admits a completion by small components, and then only through smoothing kernels. The class constants $c[j,1,b]$ are data of this kind, constant on each bin of $r = 1 - \tau$. Since Conjecture~\ref{conj:wall} allows any finite partition into bins, it covers such data when they are step functions of $\tau$ with finitely many steps, and no others. Two richer families of this kind were tried, at $\eta = 1/36$ on a grid of step $1/72$ unless stated otherwise. Adding free functions $G_j(\tau)$ on vectors with exactly one rough component, which still see the small primes only through their number and total size, gives $-0.272$ at $\alpha' = 25/72$ for $U = 4$, $C = 50$, and $-0.410$ for $U = C = 50$, $200$, $1000$ and $5000$ (these programmes omit $G_j$ at small total exactly $1/2$ and some additional feasible high-$j$ columns, as specified in the README; these are restrictions, and the value also depends on the treatment of $E$). As functions of $\tau$ these contain the finite step functions that Conjecture~\ref{conj:wall} covers with finer bins; the programme treats $G_j$ as values at grid points, a different discretisation from the forced family with bins of one grid step reported above, which gives $-0.265$ at the same point for $U = 1000$, $0.15$ higher. For such data the rows of Lemma~\ref{rem:dead} are an integral equation of the first kind, with smoothing kernels; that is consistent with values well above $-1$, but proves nothing. Free data that also see the largest small component $\mathfrak m$, through functions $H_j(\tau, \mathfrak m)$ for $\tau \le \alpha'$ and $G_j(\tau, \mathfrak m)$ on vectors with one rough component (on some grids with omitted $G_j$ columns as above; see the README), were computed at the ten points of Table~\ref{tab:richer}, for $U = 4$ and $C = 50$, and on the grid of step $1/72$ also for $U = C = 1000$. For $\eta \le 1/36$, at the least level computed at or past $1/3$ on each grid ($1/3$ for $\eta = 1/36$, $31/90$ for $\eta = 1/45$, $17/48$ for $\eta = 1/48$; no run was made at $1/3$ for the last two), the values lie within $0.05$ of $-1$: below it, with either treatment of $E$, on the grid of step $1/108$ at $\eta = 1/36$ and at $\eta = 1/45$, and within $0.003$ of it on the other two grids. At $\eta = 1/48$ the value changes side when $E$'s constant is fixed at $0$. On the grid of step $1/72$ at $\eta = 1/36$, for $U = 4$, $C = 50$ both displayed values exceed $-1$, but the original weak-duality bound lies below $-1$; for $U = C = 1000$, fixing $E$'s constant at $0$ changes the value from below $-1$ to above it. At the next grid point the values lie between $-0.96$ and $-0.78$. By the acceptance rule above, the two values below $-1$ with either treatment of $E$ are numerical prime-free candidates of this family on their grids. We draw no conclusion from them. \begin{table}[ht] \centering\small \begin{tabular}{cccrrr} $\eta$ & grid step & $\alpha'$ & best $f(1)$ & weak-duality bound & with $E$'s constant at $0$\\ \hline $1/36$ & $1/72$ & $1/3$ & $-0.9999986^\dagger$ & $-1.0000349$ & $-0.9970$\\ $1/36$ & $1/72$ & $25/72$ & $-0.9025$ & $-0.9027$ & \\ $1/36$ & $1/108$ & $1/3$ & $-1.0139$ & $-1.0141$ & $-1.0119$\\ $1/36$ & $1/108$ & $37/108$ & $-0.9483$ & $-0.9484$ & \\ $1/45$ & $1/90$ & $31/90$ & $-1.0415$ & $-1.0417$ & $-1.0391$\\ $1/45$ & $1/90$ & $32/90$ & $-0.9578$ & $-0.9580$ & $-0.9555$\\ $1/48$ & $1/96$ & $17/48$ & $-1.0004^\dagger$ & $-1.0006$ & $-0.9981$\\ $1/48$ & $1/96$ & $35/96$ & $-0.7828$ & $-0.7831$ & \\ $1/30$ & $1/90$ & $1/3$ & $-0.8847$ & $-0.8848$ & \\ $1/27$ & $1/108$ & $1/3$ & $-0.7973$ & $-0.7974$ & \\ \end{tabular} \caption{The family whose free data also see the largest small component, for $U = 4$, $C = 50$ (floating point; the last column fixes the constant of $E$ read as two rough components at $0$, so that $T_0'x = 0$; values are rounded up and bounds down). $^\dagger$On the grid of step $1/96$, the two treatments of $E$ give values on opposite sides of $-1$. On the grid of step $1/72$, for $U = 4$, $C = 50$, both values are above $-1$, but the original weak-duality bound is below it. At the latter grid point with $U = C = 1000$, the value is $-1.0000649$, and $-0.9971$ with $E$'s constant at $0$.} \label{tab:richer} \end{table} In a discretised version of the whole problem, every factorisation shape on a grid of step $1/K$ (a partition of $K$ into parts of at least $\eta K$) carries its own value, subject to the Type~I rows at level $1/2$, with the level closed ($|v| \le K/2$, i.e.\ Type~I at moduli of size exactly $x^{1/2}$; on a grid these rows include the one that couples the vector $E$ to the other shapes with a part $K/2$, which the family programmes above do not carry), and the twisted rows up to level $\alpha$, with values between $0$ and $4$ times the baseline. On three of the four grids we computed, this problem has prime-free solutions past $1/3$; on the fourth, $K = 30$, it has one at $1/3$ and none at the next grid point. At $\eta = 1/36$ it has one at $\alpha = 13/36$ and none at $14/36$ for $K = 36$, and one at $26/72 = 13/36$ and none at $27/72$ for $K = 72$; at $\eta = 1/30$ it has one at $1/3$ and none at $11/30$ for $K = 30$, and one at $21/60$ and none at $22/60$ for $K = 60$. Each ``none'' is stated for Type~I at $|v| \le K/2$ (closed) and values at most $4$ times the baseline; both are load-bearing. For $K = 36$, with the Type~I level open ($|v| < K/2$) and values at most $4$ times the baseline, the problem has a floating-point prime-free solution at $14/36$ and a certified ``none'' at $15/36$; for $K = 60$, values up to $100$ times the baseline give one at $22/60$ (below). These comparisons do not give a uniform grid-step error bound, and the other grids were not recomputed under either change. For $K = 36$ and $K = 72$ at $\eta = 1/36$ and $\alpha = 13/36$, prime-free solutions exist exactly: each of these two programmes has a rational solution whose value at the prime is exactly $0$, certified by separately written checkers in exact integer and outward-rounded arithmetic, from data that the supplement's scripts regenerate. The other two prime-free solutions, at $\eta = 1/30$, are floating-point solutions with row residuals below $3 \cdot 10^{-15}$. Each ``none'' carries a rigorous certificate: a dual solution whose weak-duality bound over all shapes is evaluated in integer arithmetic, with a rigorous allowance for the final rounding, computed by the programme that found the dual solution; the dual vectors are not included in the supplement and the bound was not recomputed by separate code. These statements are for values at most $4$ times the baseline: with values up to $100$ times the baseline, at $\eta = 1/30$ and $K = 60$, the problem has a prime-free solution at $22/60$ as well (floating point, row residuals below $3 \cdot 10^{-16}$, values up to about $99$ times the baseline). At $(\eta, \alpha) = (1/36, 13/36)$, letting the values depend on the small parts only through their number and total, and on the other parts only through the number of rough parts, the largest rough part and the large part (a family that contains the grid analogue of \eqref{eq:free}), raises the least weight on the primes to about $0.83$ times the baseline for $K = 36$ and $0.81$ for $K = 72$. Resolving the largest small part as well lowers these to $0.41$ and $0.45$, and resolving the two largest to $0.03$ and $0.096$; with the three largest the value is exactly zero for both $K = 36$ and $K = 72$, since the exact solutions above take one value per class of this family. Whether a bounded number of largest small parts suffices as the grid is refined is open. These grid results are consistent with Conjecture~\ref{conj:wall}, and they suggest that the obstruction at $1/3$ belongs to the families rather than to the problem; neither is shown. Conjecture~\ref{conj:wall} concerns this family of constructions. It says nothing about whether Hooley's condition, with Type~I information at level $1/2$ and a small Type~II range, forces primes at any level not covered by Theorem~\ref{thm:A}; for instance, for $\nu < 1/30$ that question is open at every $\alpha \in [4/15, 1/2]$; above $1/2$ we take no position (\S\ref{sec:intro}).