\section{The shared graph comparison}\label{sec:graph} This section proves the graph reduction common to the reflected and fixed-shift families of Theorem~\ref{thm:model-comparison}. Its output is \eqref{eq:gr-comparison}--\eqref{eq:gr-comparison-term}: every error correlation is replaced by comparisons containing two independent large prime products. Sections~\ref{sec:fixed-shift-comparison} and \ref{sec:reflected-comparison} estimate those comparisons in their different frequency windows. No arithmetic form of the indicators is used here beyond \textup{(C1)--(C3)}. The labels lie in one fixed reduced class. The character frame must therefore identify restrictions differing by a known unit phase. Write \(L=\log x\). There are \(\asymp\log L\) small banks and big banks, with primes respectively in \[ [\exp(L^{1/8}),\exp(L^{1/4})],\qquad [\exp(L^{1/3}),\exp(L^{2/5})]. \] Every label is \(c\bmod L_0\); primes dividing the nonzero displacement \(k\) are removed. We use \(k=T=x\) for sums and fixed \(k=h\) for shifts. All banks are disjoint. Their laws are \(\mu_g(p)=(V_gp)^{-1}\), where \(V_g=\sum_{p\in\mathcal P_g}p^{-1}\) is bounded above and below by positive fixed constants. Banks in this section may be any active subfamilies of the banks constructed from the class in \textup{(C2)}, provided that each size has at least a fixed positive multiple of \(\log L\) banks. All constants may depend on these fixed bounds and on \(L_0\). \subsection{The character frame for progression banks} The ideal-kernel construction in \cite[Section~4]{weightedgraphs} uses primitive characters of conductor at most \(Q=L^{q_*}\). On labels \(c\bmod L_0\), this list is redundant up to phases. For example, the principal character and \(\chi_{-4}\) agree on primes \(1\bmod4\) and are opposite on primes \(3\bmod4\). The quotient below removes this redundancy while retaining its phase. Let \(\mathcal X_Q\) be the primitive characters of conductor at most \(Q\), and define \begin{equation} \eta\sim\eta' \quad\Longleftrightarrow\quad \operatorname{cond}(\eta\overline{\eta'})\mid L_0. \label{eq:gr-equivalence} \end{equation} The conductor here belongs to the primitive character inducing the quotient on a common modulus. Choose one representative from each class and denote the resulting set by \(\mathcal R_Q\). Thus \(|\mathcal R_Q|\le Q^2\). To verify the equivalence and its meaning, put \(M_0=\lcm(L_0,\operatorname{cond}\eta,\operatorname{cond}\eta')\). Reduction from \((\Z/M_0\Z)^\times\) onto \((\Z/L_0\Z)^\times\) is surjective. Its kernel consists of the units \(1\bmod L_0\). A character is trivial on that kernel exactly when it factors through the latter group. This is equivalent to its primitive conductor dividing \(L_0\). Consequently, if \(\eta\sim\eta'\) and their quotient is induced by \(\vartheta\bmod L_0\), then on a product of \(r\) labels, \[ \eta(D)=\vartheta(c)^r\eta'(D). \] Both comparison products have the same slot type and the same phase. Inequivalent characters have zero average on the kernel, and also on any fixed coset of it: translation multiplies their sum by a unit phase. Fix a nonempty set \(\Lambda\) of at most \(m\) slots per big bank, where \(m\) is fixed before the eventual graph mark length. Let \(D_\Lambda\) be the product of independent labels in these slots. Partition its logarithm into cells of width \(\tau>0\) \(I_j=[j\tau,(j+1)\tau)\), and write \(\nu_j=\Prob(\log D_\Lambda\in I_j)\). For inequivalent \(\eta,\eta'\in\mathcal R_Q\), and any prescribed \(A>0\), Siegel--Walfisz gives \begin{equation} \left|\E\big[ \1_{\log D_\Lambda\in I_j} \eta(D_\Lambda)\overline{\eta'(D_\Lambda)}\big]\right| \ll_A L^{-A}. \label{eq:gr-cell} \end{equation} Here and below the accuracy in Siegel--Walfisz is chosen after all fixed logarithmic exponents. Indeed, freeze all but one label. Its allowable values form an interval inside its original bank interval, still restricted to \(c\bmod L_0\). On a common modulus at most \(L_0Q^2\), the sum of \(\eta\overline{\eta'}\) over the appropriate reduced residue classes is zero. Siegel--Walfisz, summed over these classes, therefore gives \[ \sum_{\substack{p\le z\\p\equiv c\ (L_0)}} \eta(p)\overline{\eta'(p)} \ll_K L_0Q^2z(\log z)^{-K}. \] Partial summation bounds the reciprocal-prime sum on any subinterval of \([\exp(L^{1/3}),\exp(L^{2/5})]\) by \(O_K(L_0Q^2(L^{1/3})^{1-K})\). This is an absolute bound; no lower bound for the subinterval's length or harmonic mass is needed. Divide by \(V_g\), then average the frozen labels. Deleting target divisors changes the answer by at most \begin{equation} \sum_{\substack{p\mid k\\p\ge\exp(L^{1/3})}}\frac1p \le \frac{\log(2|k|)}{L^{1/3}}\exp(-L^{1/3}), \label{eq:gr-deletion} \end{equation} which is smaller than every fixed inverse power of \(L\). This proves \eqref{eq:gr-cell} uniformly for the actual banks. On cells with \(\nu_j\ge\xi\), define \begin{equation} K_\Lambda(D,X)= \sum_{\nu_j\ge\xi} \frac{\1_{\log D\in I_j}\1_{\log X\in I_j}}{\nu_j} \sum_{\eta\in\mathcal R_Q}\overline{\eta(D)}\eta(X). \label{eq:gr-kernel} \end{equation} Both product variables have the same law. The elementary bounds are \begin{equation} |K_\Lambda|\le Q^2\xi^{-1},\qquad \sup_X\E_D|K_\Lambda(D,X)|\le Q^2,\qquad \#\{j:\nu_j>0\}\ll_m1+(\log L)L^{2/5}/\tau. \label{eq:gr-kernel-bounds} \end{equation} The last inequality uses only the range of the logarithm. On a retained cell, the kernel is the positive frame operator \[ f\longmapsto \sum_{\eta\in\mathcal R_Q}\eta\,\langle\eta,f\rangle \] for the conditional product probability measure. Its Gram matrix has diagonal entries one, since all labels exceed \(L_0Q^2\), and off-diagonal entries \(O_A(\xi^{-1}L^{-A})\) by \eqref{eq:gr-cell}. Thus its norm is at most \(1+O_A(Q^2\xi^{-1}L^{-A})\le2\). Distinct cells are orthogonal blocks, so \begin{equation} \|K_\Lambda\|_{2\to2}\le2. \label{eq:gr-frame-norm} \end{equation} This also holds for tuple tests: conditional expectation from tuples to their product is a contraction. If \(\psi\) is a character of modulus \(r\le Q\), its primitive inducing character is equivalent to one member of \(\mathcal R_Q\). For \(r_\Lambda\) selected slots and \(\psi=\vartheta\eta\) on these units, that summand gives \(\vartheta(c)^{r_\Lambda}\eta(X)=\psi(X)\). The other summands give \begin{equation} \E_D K_\Lambda(D,X)\psi(D) =\psi(X)+O_A(Q^2\xi^{-1}L^{-A}) \label{eq:gr-reproduce} \end{equation} at every retained \(X\). This is where quotienting the frame is essential: the unquotiented frame can reproduce the same constant several times. \begin{lemma}[Progression ideal comparison]\label{lem:gr-ideal} For every prescribed ideal accuracy \(E_{\rm id}\) and fixed phase range \(|\zeta|\le L^{C_4}\), the progression banks admit the signed comparison family of \cite[Theorem~4.1]{weightedgraphs}. Its ideal norm is at most \(L^{-E_{\rm id}}\), uniformly in every real additive frequency. Its coefficients ignore small-bank labels and shared labels, free at most \(m\) slots per big bank, and have total supremum at most \(L^{C_1}\). The parameters \(m,C_1\) and all kernel powers are fixed before the eventual mark length. \end{lemma} \begin{proof} We give the comparison estimates used in the construction, including the only step where restricting the prime labels changes the proof. For a product formed from \(k_g\le m\) draws in each of at most \(C_0\log L\) disjoint banks, unique factorization gives its exact law \begin{equation} n\Prob(D_\Lambda=n) =\prod_g\frac{k_g!}{V_g^{k_g}\prod_{p\in\mathcal P_g}a_p!}, \qquad \Prob(D_\Lambda=n)\le L^{K_0}/n, \label{eq:gr-productlaw} \end{equation} where \(K_0=C_0(\log(m!)+m\log\max(1,v_-^{-1}))\) and \(V_g\ge v_->0\). The factor \(C_0\) multiplies both terms of this exponent. For a tuple test, Cauchy--Schwarz on each product fibre gives an integer coefficient sequence of squared norm at most \(L^{K_0}/U\) times the tuple squared norm on a product dyad \([U,2U)\). The elementary integer bilinear estimate \cite[Lemma~4.2]{weightedgraphs}, or \cite[Lemma~6.2]{walsh}, gives \begin{equation} \left|\sum_{u,v}a_ub_v\ee{\theta uv}\right|^2 \ll \|a\|_2^2\|b\|_2^2 (1+U/r)(V+r\log(2r)) \label{eq:gr-integer-bilinear} \end{equation} when the two supports lie in containing intervals of lengths \(O(U),O(V)\), \((a,r)=1\), and \(|\theta-a/r|\le r^{-2}\). For completeness, Cauchy--Schwarz and expansion leave \(V+\sum_{1\le j\ll U}\min(V,\|\theta j\|^{-1})\). Divide the \(j\)'s into blocks of diameter less than \(r/2\). Within a block the rational parts are \(1/r\) separated and their perturbations differ by at most \(1/(2r)\); hence the sum is \(O(V+r\log(2r))\) per block. This proves \eqref{eq:gr-integer-bilinear} and permits arbitrary subsets of the containing intervals. Apply this estimate to two nonempty selected product sets \(A,B\). On product dyads \(U,V\), use Dirichlet approximation with denominator ceiling \(\lfloor UV/S\rfloor\). When the resulting denominator exceeds \(Q\), \eqref{eq:gr-productlaw} gives the normalized bound \begin{equation} \ll L^{K_0}\big(o(L^{-A})+Q^{-1}+L^{7/5}/S\big)^{1/2}. \label{eq:gr-highden} \end{equation} There are at most \(L^{7/5}\) product dyads per variable. The same bound applies to the kernel terms by moving each kernel onto its endpoint test using \eqref{eq:gr-frame-norm}. The factor \((uv)^{i\zeta}\) separates and preserves test norms. For denominators \(r\le Q\), all products are units modulo \(r\). The multiplicative-character expansion of the rational phase has total absolute coefficient at most \(\sqrt{\varphi(r)}\). Replace each primitive character by its representative in \(\mathcal R_Q\) and apply \eqref{eq:gr-reproduce}. Grouping characters after inserting their unit phases cannot increase coefficient mass, so the reproduction error is \(O_A(Q^{5/2}\xi^{-1}L^{-A})\). Write the remaining phase as \(M(v)=\exp(i\zeta v)\ee{\beta\exp v}\). On matched cells its derivative is \(O(L^{C_4}+S)\), and the two logarithmic products move by at most \(2\tau\). By \eqref{eq:gr-kernel-bounds}, this costs \(O(Q^4\tau(S+L^{C_4}))\). The union of discarded cells has probability at most \((N_A+N_B)\xi\), where \(N_A,N_B\) count the possible cells. Cauchy--Schwarz bounds its contribution by the square root of this probability times the two test norms. These estimates prove the two-product comparison, for arbitrary \(L^2\) tuple tests, to any prescribed power \(L^{-A_0}\). There is no circular choice of precisions. One possible hierarchy is \[ \begin{split} B_0&>A_0+2(7/5)+10,\\ q_*,s_*&>2(B_0+K_0+10)+7/5,\\ \tau_*&>4q_*+\max(s_*,C_4)+B_0+10,\\ x_*&>2B_0+7/5+\tau_*+10, \end{split} \] with \(Q=L^{q_*},S=L^{s_*},\tau=L^{-\tau_*},\xi=L^{-x_*}\). Choose the accuracy in \eqref{eq:gr-cell} afterwards. All these choices are independent of the eventual mark length. Here is the signed cancellation which completes the ideal construction. Split the big banks into two blocks of comparable size and designate \(m\) probe slots per group. For nonempty probe subsets \(A,B\) in the two blocks, free their slots and give the comparison coefficient \begin{equation} -(-1)^{|A|+|B|}K_A(D_A,X_A)K_B(D_B,X_B). \label{eq:gr-signed} \end{equation} Here \(D_A,D_B\) are auxiliary-label products and \(X_A,X_B\) are the unshared endpoint products. The coefficient has no small or shared labels. There are at most \(2^{mC_0\log L}\) comparisons; by \eqref{eq:gr-kernel-bounds} their total supremum is at most \begin{equation} 1+2^{mC_0\log L}Q^4\xi^{-2}\le L^{C_1}. \label{eq:gr-coeffcost} \end{equation} Expanding both kernels also separates each coefficient into a sum of products of four bounded factors: one on each auxiliary product and one on each unshared endpoint list. The number of terms and total scalar coefficient mass are fixed powers of \(L\). We now make the exceptional projection and the cancellation explicit. Write the two big-bank blocks as \(\mathcal B_1,\mathcal B_2\), of sizes \(s_1,s_2\), and their designated probe coordinates as \(\mathcal I_1,\mathcal I_2\). For a coordinate \(i\), let \(E_i\) average its independent label. These commuting operators are orthogonal projections. Let \(Q_{\rm tail}\) average the last \(\ell\) input or target coordinates in every group, and let \(\Pi\) be within-group symmetrization on all \(M=J+\ell\) coordinates. The ideal multipliers do not use those last coordinates. For symmetric endpoint tests \(F,H\), put \(F'=Q_{\rm tail}F,\ H'=Q_{\rm tail}H\). Both averaging operations are contractions. Let \(D_{\rm sh}\) be the product of the first \(J\) coordinates, and write \[ \Phi_{\Theta,\zeta} =D_{\rm sh}^{i\zeta}\ee{\Theta D_{\rm sh}},\qquad R(F',H')=\E\big[F'\overline{H'}\Phi_{\Theta,\zeta}\big]. \] Thus \(R\) is a bilinear pairing of norm at most one. Resolve the probes in the first block of \(F'\) and the second block of \(H'\): \[ \begin{split} F'&=\sum_{S\subset\mathcal I_1}F_S,\qquad F_S=\prod_{i\in S}(I-E_i) \prod_{i\in\mathcal I_1\setminus S}E_iF',\\ H'&=\sum_{T\subset\mathcal I_2}H_T,\qquad H_T=\prod_{i\in T}(I-E_i) \prod_{i\in\mathcal I_2\setminus T}E_iH'. \end{split} \] These are orthogonal decompositions. Define \(P_1=\prod_{i\in\mathcal I_1}(I-E_i)\), \(P_2=\prod_{i\in\mathcal I_2}(I-E_i)\), so that \(F_{\rm full}=P_1F'\), \(H_{\rm full}=P_2H'\). In one group the original symmetric vector decomposes orthogonally over subsets of its \(M\) coordinates. Symmetry makes the squared norms equal among subsets of cardinality \(j\). Requiring all \(m\) probes to occur and all last \(\ell\) slots to be absent therefore retains exactly the fraction \[ \frac{\binom{M-m-\ell}{j-m}}{\binom Mj} =\frac{(j)_m(M-j)_\ell}{(M)_{m+\ell}} \le 2(j/M)^m(1-j/M)^\ell \le2\ell^\ell m^{-\ell}=:\theta_m \] when \(M\ge(m+\ell)(m+\ell-1)\); impossible cardinalities give zero. For a vector which is not a tensor product across groups, first resolve by the tuple of cardinalities in the groups of a block. Independent within-group permutations make all subset tuples of those cardinalities have equal squared norms. The retained fraction is consequently the product of the displayed fractions. Summing over the cardinality tuples proves the operator estimates \begin{equation}\label{eq:gr-blockprojection} \|P_1Q_{\rm tail}\Pi\|\le\theta_m^{s_1/2},\qquad \|P_2Q_{\rm tail}\Pi\|\le\theta_m^{s_2/2}. \end{equation} Averaging tails outside the block is a further contraction. We use \(\ell=1\), and each \(s_i\ge c_0\log L/3\). For a component pair \(F_S,H_T\), put \(U=\mathcal I_1\setminus S\), \(V=\mathcal I_2\setminus T\). A comparison freeing \(A\subset\mathcal I_1\), \(B\subset\mathcal I_2\) vanishes unless \(A\subset U,B\subset V\). Indeed, if a freed input coordinate belongs to \(S\), it is mean zero, absent from the target, and absent from the comparison multiplier: integrating it gives zero. The corresponding statement holds for a freed target coordinate belonging to \(T\). This uses the orientation in \eqref{eq:gr-signed}: the \(A\) endpoint labels are from the target and the \(B\) endpoint labels are from the input. If \(A\subset U,B\subset V\), condition on all shared coordinates outside \(A\cup B\), with product \(C\). The input test is independent of the freed \(A\) labels and the target test is independent of the freed \(B\) labels. The two-product kernel comparison established above therefore replaces the unsigned comparison by \(R(F_S,H_T)\), with error \[ L^{-A_0}\|F_S\|_2\|H_T\|_2. \] Its frequency is \(\Theta C\); uniformity for every real frequency is essential. Cauchy--Schwarz over the conditioned coordinates gives exactly the displayed unconditional norms. For nonempty \(U,V\), the total unsigned subset sign is \[ \sum_{\substack{\varnothing\ne A\subset U\\ \varnothing\ne B\subset V}} (-1)^{|A|+|B|} =\left(\sum_{\varnothing\ne A\subset U}(-1)^{|A|}\right) \left(\sum_{\varnothing\ne B\subset V}(-1)^{|B|}\right)=1. \] The initial minus sign in \eqref{eq:gr-signed} cancels the raw pairing. If either missing set is empty, no comparison survives. All exceptional raw component pairs together have the exact identity \[ R_{\rm exc}(F,H)= R(P_1F',H')+R((I-P_1)F',P_2H'). \] In operator notation, if \(M_\Phi\) multiplies by \(\Phi_{\Theta,\zeta}\), their operator on the symmetric space is \[ T_{\rm exc}= \Pi Q_{\rm tail} \big[M_\Phi P_1+P_2M_\Phi(I-P_1)\big] Q_{\rm tail}\Pi . \] Consequently \[ \|T_{\rm exc}\| \le\theta_m^{s_1/2}+\theta_m^{s_2/2}. \] This identity sums the exceptional components before taking an absolute value. It is the reason their bound has no subset-count or comparison-coefficient factor. There are at most \(3^{P}\) admissible component/free-subset triples, where \(P=m(s_1+s_2)\le mC_0\log L\): each probe is indicated, missing but not freed, or freed. Thus the full signed ideal operator satisfies \[ \|T_{\rm signed}-T_{\rm exc}\| \le3^{P}L^{-A_0}. \] Choose \(m\) so that \(\theta_m<1\) and \(-c_0\log\theta_m/6>E_{\rm id}+3\), and then choose \(A_0>E_{\rm id}+mC_0\log3+10\). The kernel precisions above and the exponent \(C_1\) in \eqref{eq:gr-coeffcost} are chosen after these choices. Only afterwards require \(J\ge m\) and \(J+\ell\ge(m+\ell)(m+\ell-1)\); increasing this fixed \(J\) to meet the physical root-moment theorem changes no kernel parameter or coefficient exponent. The resulting ideal norm is at most \(L^{-E_{\rm id}}\). This proves the asserted ideal norm and coefficient conditions. \end{proof} \subsection{A large displacement in the physical graph} \begin{lemma}[Large-target root moment and pairing]\label{lem:gr-large} The root moment and endpoint pairing from \cite[Theorem~3.5 and Corollary~3.11]{weightedgraphs} hold for any nonzero integer \(k\) coprime to the bank primes, with the same logarithmic accuracies. In particular they hold for \(k=T\asymp x\) after target-prime deletion. The first endpoint is required to be independent of its marks, exactly as in the imported pairing. \end{lemma} \begin{proof} Fix the labels, patterns, and permutations of a closed walk of length \(2R\), where \(R=\lceil L^{0.52}\rceil\). Its position offsets are \(ks_i\). Return to the initial integer is equivalent to \(s_{2R}=0\), independent of the nonzero \(k\). The root dependence consists of divisibility and damping conditions at the offsets \(ks_i\). The independent-residue replacement \cite[Lemma~3.6]{weightedgraphs} is uniform in the numerical values of these offsets. Indeed its active modulus has logarithm \(O_M(R(\log L)L^{2/5})=o(L)\). The Bonferroni truncation has the same modulus and term count, and each interval CRT error is \(O(1)\), independently of the specified residues. Once the deterministic damping from active primes is factored out, the error is relative and can be summed with absolute path coefficients using the imported absolute row bound. In the resulting independent residue space, replace the root residue \(z_p\) by \(k^{-1}z_p\bmod p\) for every bank prime \(p\). Every fixed-path divisibility and damping expectation is then exactly the expectation for \(k=1\). The remaining signed coefficients are unchanged. Thus the \(k=1\) root moment gives the identical bound for this \(k\). Negative root intervals follow either from the same CRT count or from \(n\mapsto-n,\ k\mapsto-k\). It remains to avoid a large factor in the endpoint pairing. Let \(D_{\max}\) bound the edge products, so that \(\log D_{\max}=O_M((\log L)L^{2/5})\). Partition the first-endpoint interval \(I\) into blocks of length \(\lceil |k|D_{\max}\rceil\), as in the source pairing, and refine each block by residue modulo \(|k|\). The operator preserves these residues. Each refined block has \(O(D_{\max}+1)\) integer positions. For \(B=\mathcal A_k\mathcal A_k^*\), the matrix \((\langle u_n,B^Ru_m\rangle)\) on one such block is positive semidefinite. Its largest eigenvalue is at most its trace. Here \(u_n\) is the unnormalized vector equal to one on all marks at \(n\). Mark independence of the first endpoint \(f_n\) gives \[ \langle f_j,B^Rf_j\rangle \le \Big(\sum_{n\in I_j}|f_n|^2\Big) \sum_{n\in I_j}\langle u_n,B^Ru_n\rangle. \] Spectral H\"older and Cauchy--Schwarz give the source block estimate; its only dimension factor is now \begin{equation} \big(O(D_{\max})\sup|f_n|^2\big)^{1/(2R)} \le \exp\!\left(O_M\!\left( \frac{(\log L)L^{2/5}+\sqrt L}{L^{0.52}}\right)\right)=O(1). \label{eq:gr-block} \end{equation} Second-endpoint neighborhoods have bounded overlap within a residue class and are disjoint between residue classes. H\"older over blocks uses only the sum of root traces on the whole original interval, bounded by the root theorem. No CRT estimate on an individual residue class modulo \(k\) is used. This proves the pairing uniformly in \(k\). \end{proof} For bounded invariant endpoint functions, the list moments needed here do not create a mark-dependent logarithmic loss. For a fixed ordered distinct list with product \(P\), divisibility gives \(O(X/P)\) integers in an interval of length \(O(X)\) contained in \([-CX,CX]\). For fixed mark length, \(P=x^{o(1)}\). After summing the list weights, \begin{equation} \sum_{\text{lists}}\frac1P\prod_gV_g^{-M}\le1. \label{eq:gr-list-moments} \end{equation} Thus the endpoint \(r\)-moments are \(O_r(X)\). Condition \textup{(C1)} bounds every fixed-depth model without a depth-dependent logarithmic exponent. Thus this observation is valid before the final depth is chosen. \subsection{Shared labels and the two comparison geometries} The following consequence of the pairing applies to both geometries. For a vector \(\mathbf t=(t_g)\) and a bounded function \(b\) on its ordered distinct lists, put \begin{equation} W_{\mathbf t}^b(n)= \prod_g\big(\1_{\omega_g(n)=t_g}V_g^{-t_g}\big) \sum_{\substack{\mathbf p\text{ a }\mathbf t\text{-list}\\ p_{g,j}\mid n}} b(\mathbf p). \label{eq:gr-listweight} \end{equation} Here \(\omega_g(n)\) counts distinct bank-prime divisors of \(|n|\). For fixed \(t_g\le m+1\), these weights are bounded by a fixed power of \(L\). In particular \(W_{\mathbf1}^1=\Omega_{\mathcal P}\). \begin{proposition}[Graph-to-arithmetic reduction] \label{prop:gr-arithmetic-reduction} Assume the imported graph inputs of Section~\ref{sec:inputs}. Use the progression banks of this section, with primes dividing \(k\) deleted, and take \(k=x\) for reflected sums or fixed \(k=h\ne0\) for shifts. Let \(\alpha,\beta\) be functions of the integer endpoint, of modulus at most one, invariant under repeated multiplication by every bank prime whenever both arguments are in the endpoint ranges. For sums use the negative first half-line and positive second half-line, extending the first function by absolute value; for shifts use positive endpoints. The pullback of the first endpoint to physical states must be constant on its mark lists. Take \(\ell=1\), \(M=J+1\), and a smooth \(\Psi\) supported in a fixed compact subset of \((0,\infty)\). For every fixed accuracy \(A>0\), Lemmas~\ref{lem:gr-ideal} and \ref{lem:gr-large} give fixed \(m,C\) and a sufficiently large fixed \(J\) for which \begin{equation} C_0+\sum_\nu c_\nu C_\nu=O_A(L^{-A}),\qquad \#\{\nu\}+\sum_\nu|c_\nu|\le L^{C}, \label{eq:gr-comparison} \end{equation} where \[ C_0=\sum_{w\ne0}\frac{\Psi(|w|/x)}{|w|} \alpha(w)\beta(w+k) W_{\mathbf1}^1(w)W_{\mathbf1}^1(w+k) \] and each comparison has the form \begin{equation} \begin{split} C_\nu=\E_{D_1,D_2}a_\nu(D_1)b_\nu(D_2) \sum_{w\ne0}\frac{\Psi(|w|/(xD_1D_2))}{|w|} \alpha(w)\beta(w+kD_1D_2)\\ {}\times W_{\mathbf t}^{u_\nu}(w) W_{\mathbf t}^{v_\nu}(w+kD_1D_2). \end{split} \label{eq:gr-comparison-term} \end{equation} The following restrictions and normalizations are part of the conclusion. \begin{enumerate} \item The physical pairing is restricted to \(\omega_g(n)=\omega_g(n')=M\) in every bank. Its state mass is \(\prod_gV_g^{-M}\), its transition row normalization is \(\prod_gV_g^{-t_g}\), and hence its joint endpoint normalization in bank \(g\) is \(V_g^{-M-t_g}\). After shared labels are removed, the exact counts are \(\omega_g(w)=\omega_g(w+kD_1D_2)=t_g\), as imposed by \eqref{eq:gr-listweight}, where \(t_g=1+\#\{\text{freed slots in }g\}\le m+1\). In every small bank \(t_g=1\). \item Each comparison frees nonempty probe sets in two disjoint blocks of big banks. The products \(D_1,D_2\) use these respective sets, with at most \(m\) draws per bank. All auxiliary draws are independent with laws \(\mu_g\), including between the two products; repetitions among auxiliary draws are allowed. No auxiliary-label exclusion from endpoint lists remains in the displayed comparison. \item Before separation, a signed pattern coefficient may depend on the freed big-bank auxiliary labels and the unshared big-bank endpoint labels specified in \eqref{eq:gr-signed}. It uses no small-bank labels, shared labels, or distinguished last endpoint slots. After separation, \(a_\nu\) depends only on \(D_1\), \(b_\nu\) only on \(D_2\), and \(u_\nu,v_\nu\) only on the ordered unshared big-bank labels at their respective endpoints. Each of these four factors has modulus at most one. In particular, both list coefficients ignore every small-bank label. \item The parameters \(m,C\), the comparison family and its scalar coefficient bounds are fixed before the final mark length \(J\) and the final model depth. Increasing \(J\) supplies shared slots without changing them. For endpoints satisfying \textup{(C1)}, no logarithmic loss exponent depends on the model depth; the threshold in \(x\) may depend on that depth and on \(J\). \end{enumerate} \end{proposition} \begin{proof} The signed coefficients are those of Lemma~\ref{lem:gr-ideal}; their allowed label dependence and four-factor expansion were proved in \eqref{eq:gr-signed}--\eqref{eq:gr-coeffcost}. Each auxiliary factor depends only on its numerical product: \eqref{eq:gr-kernel} uses only that product's log cell and character. Thus the expectation over ordered draws pushes forward to the product law \eqref{eq:gr-productlaw}, including its repetition multiplicities. The first endpoint is mark independent, as required by the imported pairing, while all pattern-dependent mark factors remain in the operator. A pattern has \(M-1\) dilation slots per bank, some shared and some freed, and one distinguished unshared endpoint slot. Write its edge product as \(D=D_RD_C\), with shared product \(D_R\) and auxiliary product \(D_C\). For a source of either sign in \(|n|\in[Z,2Z)\), an originally nonzero edge weight satisfies \(\Psi(|n|/(xD))\ne0\), hence \(D\ll Z/x\). Since \(|k|\ll x\), this implies \(|n+kD|\ll Z\). Insert a second-endpoint cutoff on \([-C_\Psi Z,C_\Psi Z]\) before Mellin inversion. It equals one on every contributing edge. After inversion retain this same cutoff. An edge originally having zero \(\Psi\) still has zero full Mellin integral, so the insertion does not change the identity. No cutoff involving shared-label products has been introduced. Mellin inversion of \(\Psi\), with the harmonic weight \(1/|n|\), now applies the pairing of Lemma~\ref{lem:gr-large} on each of \(O(L)\) source dyads. Its tails are bounded by the absolute row/column estimate, its main part by the small ideal norm, and its endpoint norms by \eqref{eq:gr-list-moments}. Compressing the physical operator to exact-count vertices \(\omega_g(n)=\omega_g(n')=M\) is legitimate: the corresponding projection depends only on the integer, preserves mark independence, and sets both damping factors equal to one. For completeness, put \(n=D_Rw\), \(n'=D_Rw'\), so \(w'=w+kD_C\), and put \(t_g=1+\#\{\text{freed slots in }g\}\). Invariance and harmonic scaling give \[ \alpha(D_Rw)=\alpha(w),\quad \beta(D_Rw')=\beta(w'),\quad \frac{\Psi(|n|/(xD))}{|n|} =\frac1{D_R}\frac{\Psi(|w|/(xD_C))}{|w|}. \] The graph normalization at a shared prime combines with \(1/D_R\) to give its probability law \(\mu_g\). More explicitly, with \((D_R)_g\) its product in bank \(g\), \begin{equation} \frac{V_g^{-M-t_g}}{(D_R)_g} =V_g^{-2t_g} \prod_{\text{shared slots in }g}\mu_g(p). \label{eq:gr-sharednormal} \end{equation} For fixed reduced endpoints and fixed auxiliary labels, the shared-label probability is \begin{equation} \prod_g\1_{\omega_g(w)=\omega_g(w')=t_g} +O(L^{C_M}e^{-L^{1/8}}). \label{eq:gr-sharedprob} \end{equation} Indeed there are only \(O_M(\log L)\) draws, each atom is \(O(e^{-L^{1/8}})\), and \(ww'D_C\) has \(O(L)\) prime divisors. If the reduced counts are \(t_g\), distinct draws avoiding those divisors reconstruct the exact graph edge. If a reduced count is larger, a valid reconstruction must choose a shared prime already dividing that endpoint, an event of the displayed small probability. This proves \eqref{eq:gr-sharedprob}, also for negative \(w\). Summing its absolute error costs only fixed logarithmic powers, by bounded endpoints and the harmonic weight. The restrictions excluding auxiliary labels from endpoint lists can then be removed: for an auxiliary prime \(p\mid D_C\), \(p\mid w\) if and only if \(p\mid w+kD_C\); these positions have density \(O(1/p)\) in the containing interval. Every auxiliary prime is at least \(\exp(L^{1/3})\). An unshared prime common to the two endpoints divides \(kD_C\); bank-prime deletion implies it divides \(D_C\), so the same estimate covers it. Thus these errors remain smaller than every logarithmic power. Finally expand the separated coefficients from \eqref{eq:gr-signed}. Formula \eqref{eq:gr-sharednormal} yields exactly \eqref{eq:gr-listweight}, while \(D_C=D_1D_2\). The raw pattern has \(D_C=1,\mathbf t=\mathbf1\). This proves \eqref{eq:gr-comparison}--\eqref{eq:gr-comparison-term} with their stated independence properties. The kernel parameters and coefficient exponents were fixed before \(J\) in Lemma~\ref{lem:gr-ideal}. Endpoint boundedness and \eqref{eq:gr-list-moments} pay the list norms without a depth-dependent logarithmic power. Thus the final model depth can be chosen later to meet the one-variable accuracy, changing only the threshold for boundedness in \textup{(C1)}. \end{proof}