\section{Replacing indicators by arithmetic models}\label{sec:comparison} We isolate the prime invariance, one-variable means and sparse sieve bounds needed to replace an indicator by its arithmetic model. Fix an even integer \(L_0\ge2\) and a reduced class \(c\in(\Z/L_0\Z)^\times\). At the outer scale \(X\), write \[ L=\log X,\qquad y=\exp(L^{2/5}),\qquad \mathcal P_y=\{p:p\le y\},\qquad E=\exp(\sqrt L). \] For a set of primes \(\mathcal P\), let \(n_{\mathcal P}\) be the part of \(n\) supported on \(\mathcal P\), including its full valuations. A rough integer has no prime factor in \(\mathcal P_y\). There are two families of lines in the theorem: \begin{align*} \mathcal L_X^{\rm sum}&=\{(u,v)\in\Z_{>0}^2:Au+Bv=X\}, &t(u,v)&=Au, &J_X&=[0,X],\\ \mathcal L_X^{\rm shift}&=\{(u,u+h):u\in\Z,\ X\le u\le2X\}, &t(u,v)&=u, &J_X&=[X,2X]. \end{align*} In the first family \(X\) is an integer target and \(A,B\) are fixed coprime positive integers supported on primes dividing \(L_0\). In the second, \(h\ne0\) is a fixed integer and \(X\) tends to infinity; both coordinates are then positive. Set \(k=X\) in the sum family and \(k=h\) in the shift family. Dependence on the fixed data, including \(h\) in the second family, is permitted. Let \(0\le f_i\le1\), \(i=1,2\), be fixed functions on the positive integers. For every sufficiently large fixed integer \(K\), let \(\mathcal D_{i,X,K}\) be an arithmetic model, and write \(\xi_{i,X,K}=f_i-\mathcal D_{i,X,K}\). We often suppress \(X,K\). The entire model, its cutoff \(y\), and its small-prime records stay tied to this outer scale during all subsequent dilations. Choose a positive comparison scale \(\mathcal S_X\) satisfying \begin{equation}\label{eq:comparison-scale} XL^{-\kappa_0}\le\mathcal S_X\le XL^{\kappa_0} \end{equation} for a fixed \(\kappa_0\ge0\). Arithmetic factors may be included in this scale. The hypotheses are as follows. \begin{description} \item[\textup{(C1)} Bounded models.] For every fixed \(K\), sufficiently large \(X\) gives \(0\le \mathcal D_{i,X,K}(n)\le1\) on \(1\le n\le X^3\). The bound one is independent of the model depth; its threshold may depend on \(K\). \item[\textup{(C2)} A common prime class.] For every prime \(p\le y\) with \(p\equiv c\pmod{L_0}\), \[ f_i(pn)=f_i(n),\qquad \mathcal D_{i,X,K}(pn)=\mathcal D_{i,X,K}(n) \] whenever the arguments are in the range under consideration. This includes repeated multiplication by \(p\). A residue filter is part of \(f_i\) and must satisfy the same equality. \item[\textup{(C3)} Centered one-variable rough means.] For every fixed \(B_0,\mathcal D_0>0\), there is a depth \(K_0\) such that every fixed \(K\ge K_0\) has the following property. Uniformly for \(\mathcal P_y\)-supported integers \(r\le E\), \(X^{1/2}\le Y\le X^2\), bounded periodic functions \(a(v)\) of period at most \(L^{B_0}\), and smooth functions \(W\) supported in a fixed compact subinterval of \((0,\infty)\), \begin{equation}\label{eq:comparison-rough-means} \left|\sum_{v\ {\mathrm{rough}}} \xi_{i,X,K}(rv)a(v)W(rv/Y)\right| \ll_{B_0,\mathcal D_0,K}\frac Yr L^{-\mathcal D_0}. \end{equation} Here \(|a|\le1\), and every fixed logarithmic derivative of \(W\) used in the estimate is at most a fixed power of \(L^{B_0}\). The hypothesis permits Mellin factors with polynomial-logarithmic frequency, by including them in \(W\). Its constants may depend on the chosen derivative orders. Every surviving character branch, including a negative or complex pole parameter, must be retained when verifying this estimate. \item[\textup{(C4)} A positive sparse record envelope.] Take disjoint prime banks in the two ranges \[ [e^{L^{1/8}},e^{L^{1/4}}],\qquad [e^{L^{1/3}},e^{L^{2/5}}], \] each bank consisting of the primes in an interval which are \(c\pmod{L_0}\), after deleting primes dividing \(k\). Assume each bank has reciprocal mass between \(1/2\) and \(2\), and that the number in either range is at most \((\log L)/(16\varphi(L_0))\). Write \(V_g=\sum_{p\in\mathcal P_g}p^{-1}\) for its reciprocal mass. Let \(\mathcal B\) be their union. Define the independent joint record law \(\mu_{\mathcal B}=\prod_{p\in\mathcal B}\mu_p\) by \begin{equation}\label{eq:bank-law} \mu_p(0,0)=1-2/p,\qquad \mu_p(e,0)=\mu_p(0,e)=(1-1/p)p^{-e}\quad(e\ge1), \end{equation} and zero mass on two positive valuations. For every fixed \(B_0>0\), nonnegative record function \(F(r,s)\le L^{B_0}\), interval \(J\subset J_X\), and \(w_i\in\{f_i,\mathcal D_{i,X,K}\}\), require \begin{align} &\sum_{\substack{(u,v)\in\mathcal L_X\\t(u,v)\in J}} F(u_{\mathcal B},v_{\mathcal B})w_1(u)w_2(v) \ll_{B_0,K}\frac{|J|}{X}\mathcal S_X \E_{\mu_{\mathcal B}}F\notag\\ &\hspace{34mm} +X^{1/2}L^{C_{B_0}} +XL^{C_{B_0}}\exp(-c_0L^{1/10}). \label{eq:comparison-envelope} \end{align} The positive constant \(c_0\) and exponent \(C_{B_0}\) are independent of the final fixed model depth. Constants and thresholds may depend on that depth. Uniformity includes the bank intervals, the exact valuations in the record, and the interval endpoints. \end{description} For a bank \(g\), let \(\omega_g(n)\) count its distinct prime divisors, and define \[ I_g(u,v)=\1_{\omega_g(u)=\omega_g(v)=1},\qquad \Omega_T(n)=\prod_{g\in T}V_g^{-1}\1_{\omega_g(n)=1}. \] Here \(T\) is a set of bank indices. We also write \(\Omega_{\mathcal P}\) when \(\mathcal P\) denotes the union of the active banks. The joint valuation law above is the ordinary CRT law on either line: every bank prime avoids \(k\) and is a unit at the fixed coefficients. Condition (C3) is a statement about one variable on a progression; it allows arbitrary retained small-prime records, but no arbitrary weight in the rough variable. Condition (C4) is only an upper bound with nonnegative weights. In the applications it follows from CRT root counts and an upper sieve, with a power-smaller endpoint error. Neither condition assumes centered binary cancellation or a major-arc energy estimate. In particular, choosing a model with the right average density alone does not verify (C3). \begin{theorem}[Arithmetic model comparison]\label{thm:model-comparison} Assume the zero-free and weighted-dilation inputs of Section~\ref{sec:inputs}, and conditions \textup{(C1)--(C4)}. Put \(\delta=1/(8192\varphi(L_0))\). For every sufficiently deep fixed \(K\), uniformly for intervals \(J\subset J_X\), \begin{equation}\label{eq:model-comparison-conclusion} \sum_{\substack{(u,v)\in\mathcal L_X\\t(u,v)\in J}} \bigl(f_1(u)f_2(v)-\mathcal D_{1,X,K}(u)\mathcal D_{2,X,K}(v)\bigr) \ll_K\mathcal S_XL^{-\delta}. \end{equation} For the sum family the estimate is uniform in the integer target and its prime valuations. For the shift family the integer \(h\) is fixed, and the constant may depend on it. The two families have different major-arc geometries; no uniformity in the intervening range of growing sublinear shifts is asserted. \end{theorem} The proof consists of the shared graph reduction (Section~\ref{sec:graph}), the fixed-shift and reflected estimates (Sections~\ref{sec:fixed-shift-comparison}--\ref{sec:reflected-comparison}), and reconstruction (Section~\ref{sec:assembly}). Subgroup pairs satisfy (C2) in the identity class; a self-coset uses a class in that coset. Arbitrary prescriptions may have no common class. The final asymptotic saving must also pay for evaluating the model.