\section{Reconstruction and quantitative comparison}\label{sec:assembly} The results above require a positive logarithmic saving. We now record the numerical choices and prove that every approximation fits them. For the fixed subgroup theorem one may take \(\delta=1/(8192\varphi(L_0))\). In particular \(\delta_{12}=1/32768\), the coprime \(C_{12}\) saving is \(\delta_{12}/2=1/65536\), and the prime-\(3\pmod4\) self-sum saving is \(\delta_-=1/16384\). These choices concern the semigroup asymptotics; the ordinary quadratic-indicator statement has the remainder stated in Section~\ref{sec:quadratic-pairs}. We assemble the two frequency estimates into Theorem~\ref{thm:model-comparison} at the sparse scale in (C4). Write \(\phi=\varphi(L_0)\), \(L=\log X\), \(\delta_0=1/(8192\phi)\), and use \eqref{eq:comparison-scale}. \subsection{An explicit discarded-bank exponent} Primes in the reduced bank class \(c\bmod L_0\) have reciprocal mass \((1/(8\phi)+o(1))\log L\) in the small range and \((1/(15\phi)+o(1))\log L\) in the big range. Choose \[ b_X=\left\lfloor\frac{\log L}{32\phi}\right\rfloor \] disjoint banks of each size, greedily giving each reciprocal mass \(V_g=1+o(1)\). The overshoot per bank is at most the largest reciprocal prime. Delete bank primes dividing the displacement. For a reflected target their total reciprocal mass is \(O(L/(P_{\min}\log P_{\min}))\); for a fixed shift the deletion is eventually empty. Thus the total mass error is smaller than every fixed inverse logarithmic power, uniformly in the reflected target. The independent law \eqref{eq:bank-law} gives \begin{equation}\label{eq:success} \E I_g=\prod_{p\in\mathcal P_g}(1-2/p) \sum_{\substack{p,q\in\mathcal P_g\\p\ne q}} \frac1{(p-2)(q-2)} =V_g^2e^{-2V_g}+O\left(\sum_{p\in\mathcal P_g}p^{-2}\right) \longrightarrow e^{-2}>1/8. \end{equation} Banks are independent in this law. Chernoff's inequality bounds the probability of fewer than \(b_X/16\) successful banks in either range by \begin{equation}\label{eq:rare-budget} 2e^{-b_X/64}\ll L^{-1/(4096\phi)}=L^{-2\delta_0}, \end{equation} since \(b_X\ge\log L/(64\phi)\) eventually. Apply (C4) to this bad-bank indicator and use \(|f_i-\mathcal D_i|\le f_i+\mathcal D_i\). The discarded centered correlation is \(O(\mathcal S_XL^{-2\delta_0})\), plus its power-smaller and exponential endpoint errors. The reconstruction is exact. For \(b\) banks, the Boolean function asserting at least \(k\) successes has expansion \(\sum_Tc_T\prod_{g\in T}I_g\), where \(c_T=0\) for \(|T|L^3}|\widehat F_\epsilon(\zeta)|\,d\zeta \ll_j L^{-(3-\delta_0)(j-1)}. \end{equation} The external phase range stays at \(L^3\); its tail derivative order is chosen last. On a contributing reflected edge \(n<0\), \(n'=n+XD>0\), put \(t=|n|/(XD)\). Then \begin{equation}\label{eq:two-sided-localization} \frac{n'}{|n|}=\frac{1-t}{t} \in\left[\frac\epsilon{1-\epsilon}, \frac{1-\epsilon}{\epsilon}\right]. \end{equation} For a source dyad \(|n|\asymp Y\), insert a target cutoff on \([c\epsilon Y,C\epsilon^{-1}Y]\), equal to one on that range, \emph{before} Mellin inversion, and retain it in each pairing. Removed edges originally had zero full Mellin integral. Shared-label division preserves the ratio. There are \(O(\log L)\) target dyads \(Y'\), each with uniform rescaled logarithmic derivatives. Their exact-list endpoint moments are \(O(Y)\), \(O(Y')\), by counting multiples of a fixed list product. This introduces no exponent depending on the eventual mark length or density depth. Pairing costs at most \(\sqrt{Y'/Y}\ll\epsilon^{-1/2}\). The Mellin norm and dyadic count together cost at most \(L^{5\delta_0/2}\); reserve \(L^{8\delta_0}\) for all localization. For comparison scale \(Y\asymp XD_1D_2\), use the actual target dyad \(c\epsilon Y\le Y'\le C\epsilon^{-1}Y\). The reflected major window has width \(L^{B_{\rm out}}/Y\), so \(|\beta|Y'\ll L^{B_{\rm out}+\delta_0}\). Proposition~\ref{prop:gr-energy} applies at this enlarged logarithmic window exponent. Its integrated rescaling costs at most \(\epsilon^{-2}\), within the reserve. At the smallest dyad, \begin{equation}\label{eq:smallest-dyad} \frac{Y'}{qP}\gg e^{cL^{1/3}-CL^{1/4}}L^{-B_{\rm out}-\delta_0}\gg_A L^A. \end{equation} Removing a gcd divisor scales length and modulus equally, so this bound holds in every target profile. For a fixed shift the two endpoints are comparable on the original support; the cutoff and pairing cost is no larger. Its wider major windows are handled by Proposition~\ref{prop:fixed-comparison-energy}, not by extending \eqref{eq:gr-energy} beyond its proved range. \subsection{Precision order and completion} Let \(Q_0=\lceil\kappa_0\rceil+30\). Request normalized graph-pairing error \(L^{-(Q_0+3)}\) per source dyad. There are at most \(L^2\) such dyads. Fix root and ideal accuracies for this request, then the probe complexity and eventual mark length, keeping the phase range \(L^3\). Only after this family is fixed let \(C_{\rm cmp}\) dominate its coefficient and dyadic costs, and \(C_w\) its endpoint pointwise exponent. These exponents do not depend on the density depth. If \(m_*\) bounds free draws per group, the dyadic product laws obey \[ \nu_i(n)\le L^{K_*}/n,\qquad K_*=\frac{\log(m_* !)+m_*\log2}{16\phi}. \] Their mass is at most one, so their squared norm at length \(T_i\) is at most \(L^{K_*}/T_i\). In \eqref{eq:gr-multiplier} take \(C_{\rm prod}=2K_*+2\), and choose \begin{align} A_{\min}&>C_{\rm cmp}+2C_w+\kappa_0+50+8\delta_0,\notag\\ B_{\rm out}&>2A_{\min}+C_{\rm prod}+4.\label{eq:outer-budget} \end{align} Dirichlet approximation with denominator ceiling \(T_1T_2/L^{B_{\rm out}-1}\) gives multiplier \(O(L^{-A_{\min}})\) off the outer grids. Their denominators are \(Xr\) in the reflected case and \(|h|r\) for a fixed shift, with \(r\le L^{B_{\rm out}}\). The middle terms in the multiplier bound are superlogarithmically small because \(T_i\ge\exp(L^{1/3})\). Request integrated major-energy saving \(T_E>2(C_{\rm cmp}+C_w+\kappa_0+50+8\delta_0)\). In the reflected case enlarge the residue-energy accuracy to pay both arc widths and the number of outer denominators. Exact discrete Parseval accounts for all target-grid points. In the fixed-shift case use the low/high Dirichlet-frequency proof at that accuracy, choosing its smoothing and low-frequency threshold after the fixed comparison costs. The cofactor-length margin pays its mean-value bound. Both major contributions are then at most \(L^{-T_E/2+C_w+C_{\rm cmp}+8\delta_0}\). The minor bound has the same ample margin. Choose the model depth last, using (C3) at the largest period, smoothness and precision exponents now required. Its uniform-in-depth pointwise and sieve exponents in (C1) and (C4) make this order noncircular. Choose the external Mellin tail derivative order last as well. Nothing changes the earlier ideal family or phase range. \begin{proof}[Completion of Theorem~\ref{thm:model-comparison}] After localization, source-dyad summation and reconstruction, the graph and comparison errors are \[ O(XL^{-Q_0+8\delta_0})=o(\mathcal S_XL^{-\delta_0}). \] The rare-bank term is \(O(\mathcal S_XL^{-2\delta_0})\). All record, square and CRT tails are exponential or power-smaller than \(\mathcal S_XL^{-\delta_0}\). For each ordered endpoint pair use \[ f_1(x)f_2(y)-\mathcal D_1(x)\mathcal D_2(y) =(f_1(x)-\mathcal D_1(x))f_2(y)+\mathcal D_1(x)(f_2(y)-\mathcal D_2(y)). \] The banked comparison bounds each centered term to the required precision. The argument is symmetric in the endpoints; for a fixed shift, interchange them and replace \(h\) by \(-h\). To restrict to an arbitrary interval, smooth its endpoints at width \(\epsilon X\); in the reflected case first remove the two extreme strips. At most four strips of total length \(O(\epsilon X)\) are lost. Condition (C4), applied to each nonnegative pair weight, bounds their total contribution by \(O(\epsilon\mathcal S_X+X^{1/2}L^C+XL^Ce^{-cL^{1/10}})\). For an interval shorter than this width use that upper bound directly. Otherwise total variation remains \(O(1)\), and a transition at \(t_0\) has logarithmic derivative norm at most \(O_j(t_0^j\epsilon^{1-j})\). The Mellin and localization budgets are therefore uniform in both endpoints. Since \(\epsilon=L^{-\delta_0}\), every error has the asserted size. This proves the comparison theorem in both regimes. \end{proof}