% !TeX program = pdflatex \documentclass[11pt,a4paper]{article} \usepackage[margin=2.2cm]{geometry} \usepackage{amsmath,amssymb} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage{microtype} \IfFileExists{xurl.sty}{\usepackage{xurl}}{} \usepackage[colorlinks=true,allcolors=blue]{hyperref} \usepackage{bookmark} \usepackage{longtable,booktabs,array,calc} \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}} \setcounter{secnumdepth}{-\maxdimen} \setlength{\emergencystretch}{3em} \allowdisplaybreaks \hypersetup{pdftitle={Stationary prediction from finite calibration: obstructions and sharp history rates under interval catastrophes},pdfauthor={HUANG Man},pdfcreator={Pandoc; Hexagon platform compilation}} \title{Stationary prediction from finite calibration: obstructions and sharp history rates under interval catastrophes} \author{} \date{} \begin{document} \maketitle \textbf{HUANG Man}\\ Independent Researcher, Shanghai, China\\ Email: huangman@technocentrics.com\\ ORCID: https://orcid.org/0009-0003-3809-5105 \subsection{Abstract}\label{abstract} Can calibration experiments confined to a fixed spatial diameter determine stationary block probabilities on larger spatial scales? We answer this question for a one-dimensional recovery process driven by independent rows of Poisson interval catastrophes with an unknown intensity law, under a common small recovery-weighted intensity budget. For every calibration diameter, two laws have exactly the same complete local damage-word distributions and deterministic-initial experimental transcripts, even under arbitrarily many adaptive repetitions, while a sufficiently large stationary block has probabilities separated by a positive constant. The pair also matches intensity moments of orders zero, one and two and the eventual power-law parameters. For the upper stationary open phase, we then prove that all-open finite-history approximations converge uniformly over a law class and all finite targets if and only if the recovery-weighted tails vanish uniformly. Under a common intensity-moment budget of order \(2+\epsilon\), \(0<\epsilon\le1\), the worst missing-history error has matching scale \(\min\{b,B[\max(m_0,T)]^{-\epsilon}\}\), with constants depending only on a fixed small recovery budget \(b\). Here \(T\) counts retained physical rows, \(B\) is the moment budget and \(m_0\) the minimum interval length. The lower bound holds at one fixed site; the upper bound covers every finite target and the full moment class, including zero intensity and laws without regular variation. Together, these results identify the obstruction to local calibration, the exact tail condition for uniform history approximation, and the sharp deterministic depth of the all-open truncation. \textbf{Keywords:} correlated catastrophes; stationary prediction; partial identification; finite calibration; monotone recovery; history truncation. \textbf{Mathematics Subject Classification:} 60K35, 60G55, 62M20. \subsection{1. Introduction}\label{introduction} Can experiments confined to a fixed spatial diameter determine a recovery system's stationary block probabilities on larger spatial scales, even when the experiments can be repeated arbitrarily many times? If local calibration leaves an ambiguity, which tail restriction supports uniform history approximation, and how many physical rows does that approximation require? Rare long catastrophes make these questions inseparable: their frequency controls what local experiments observe, while their recovery time controls their stationary effect. We study binary states on the integer line. At each discrete time, open sites propagate to their right neighbours and independent Poisson intervals remove the resulting open sites. Damage is spatially correlated and renewed independently across time. Starting with every site open in the remote past selects the upper stationary open phase; its complement is the stationary hole field. Our information contract supplies complete fresh damage words on windows of bounded diameter, together with recovery readouts from deterministic initial fields whose causal observations satisfy the same bound. It allows adaptive choices and arbitrarily many repetitions. Calibration diameter, retained physical history and observation count are separate resources. The main results answer the three questions in sequence. First, exact local word and transcript equality can coexist with a constant gap in the stationary probability of a sufficiently large block, even after matching three intensity moments and the eventual power-law parameters (Theorem 2.1). Second, uniform vanishing of recovery-weighted tails is equivalent both to uniform single-site history approximation and to uniform approximation over all finite targets (Theorem 2.2). Third, a common moment budget yields matching worst-case history errors, with the lower bound at a fixed site and the upper bound for arbitrary finite target geometry (Theorem 2.4). Thus the information obstruction, the class criterion and the quantitative depth bound form one prediction theory for the selected phase. The common tail condition controls catastrophe mass at lengths beyond the calibrated range. Its quantitative moment version converts that control into a physical-time rate. Stationary flux and a deterministic checkpoint comparison produce an auxiliary boundary distribution; a one-source restart then controls the missing past. This route applies across a common-budget class containing zero intensity, finite support and laws without regular variation. Poisson Boolean capacity and covariogram identities describe how coverage observations encode an interval law {[}LP18, Theorems 17.3, 17.8 and 17.10{]}. Gracar, Korfhage and Mörters construct three-dimensional rotation-invariant convex-grain models with the same diameter marginals but different robustness, meaning percolation at every positive intensity {[}GKM26, Section 3.5{]}. Their ambiguity concerns dependence between grain diameters. Theorem 2.1 fixes one-dimensional interval geometry and preserves complete bounded-window damage laws and experimental transcripts before separating stationary block probabilities under recovery. Bayraktar, Lu, Maggioni, Wu and Yang study finite-state cellular automata on a finite circular graph with conditionally independent site updates determined by a local transition matrix. They obtain trajectory-based inference under nonsingularity conditions, exhibit equal one-site stationary marginals with different joint stationary laws, and give different transition matrices with the same joint stationary law {[}BLMWY26, author version v4, Examples 3.3--3.4, Assumption 4.1 and Theorem 4.4{]}. Here the unknown object is an unrestricted sequence of interval intensities. The calibration contract supplies complete joint damage words on bounded windows; the ambiguity persists even when these word laws are known exactly and the allowed experiments are repeated arbitrarily many times. Temporal coupling bounds derive memory loss from transition-kernel continuity {[}BFG99, CFF02, Theorem 4.3(iv){]}. De Santis, Laxa and Löcherbach allow unknown past symbols in coupling-from-the-past algorithms for a specified transition kernel; their lower-envelope conditions yield almost-sure termination and, under stronger conditions, integrable backward times {[}DLL25, Section 2.4, Theorems 3--6{]}. Their unknown symbols represent unresolved past states. Our unknown quantity is the catastrophe law itself, and Theorems 2.2 and 2.4 characterize deterministic physical-history approximation uniformly over a class of such laws. For spatial systems, finite ancestor clans give stationary sampling and truncation control {[}FFG02{]}, high-temperature tail conditions give finite-range approximation of Gibbs measures {[}GLO10, Theorem 1{]}, and a subcritical Kalikow backward sketch with a uniform positive lower bound on site clock rates gives exponential coupling from every initial configuration {[}GGLO13, conditions (5.1)--(5.2), Proposition 1 and Theorem 3{]}. The present results combine an unknown-law information contract with an exact class criterion and a matching lower bound; the time rate is derived from spatial catastrophe moments. Further comparisons locate the model and its assumptions. In Poisson hail, spatial job extent and service duration determine workload stability at small arrival rates {[}FKM18, Theorem 1; MW24, Theorems 1.1--1.2{]}. Here spatial extent and recovery enter a binary phase's stationary prediction error. The deterministic hole map is a two-state monotone eroder. Gács and Törmä's dependent-error stability theorem assumes an exponential bound on every finite simultaneous error set {[}GT22, Section 2{]}; the polynomially likely long error blocks admitted here violate that specific hypothesis (Appendix C). Swart, Szabó and Toninelli obtain Peierls stability bounds for monotone cellular automata with intrinsic randomness using independent space-time local-rule choices {[}SST25, (1.7)--(1.8) and Theorems 10--11{]}. Our interval atoms produce spatially correlated local-rule choices already in the first update. Recent restriction and correlation bounds treat continuous-time particle systems with bounded update diameters and unbounded dependence range, allowing phase coexistence; stationary spatial mixing is conditioned on rapid convergence from a specified initial configuration {[}JK26, assumption (R1) and Theorem 2.4{]}. We derive the physical-history rate directly for this synchronous interval model. The positive-phase and interval-insertion arguments are shared with the author's preprint \emph{An exact moment criterion for survival under renewed spatial catastrophes} {[}H26a, Version 1, Sections 3 and 5{]}, whose principal results concern survival and stationary dilute density. The companion preprint \emph{Finite-preparation response under heavy-tailed catastrophes: exact amplitudes and separation from Poisson coverage} {[}H26b, Version 1{]} develops compression, checkpoint comparisons and history estimates uniform over cutoffs of a fixed regularly varying law. Its Supplement S2.2--S2.3 supplies the one-source restart, finite-error contraction and remaining-time induction, including deterministic intensity strings. This paper uses those tools to establish exact local-law ambiguity, the common-tail equivalence and matching bounds over varying intensity sequences and all finite targets, with explicit budget and minimum-length dependence. The appendices reproduce the model-specific lemmas needed for a self-contained proof. \subsection{2. Model, information and principal results}\label{model-information-and-principal-results} \subsubsection{2.1. Damage, the selected phase and physical history}\label{damage-the-selected-phase-and-physical-history} Fix an integer \(m_0\ge2\). For each row \(t\in\mathbb Z\), origin \(z\in\mathbb Z\) and length \(\ell\ge m_0\), let \(N_{t,z,\ell}\) be independent Poisson variables of means \(\theta_\ell\ge0\). An atom covers \([z,z+\ell-1]\cap\mathbb Z\); their union on row \(t\) is \(D_t\). Assume \[ w_\ell=\frac{\ell(\ell+1)}2,\qquad \chi_\theta=\sum_{\ell\ge m_0}w_\ell\theta_\ell\le b, \qquad 03\) and \(00\), both equal \(a\ell^{-p}\) for all sufficiently large \(\ell\). \item Their selected stationary probabilities satisfy \(h_{\theta^-}(k)\le\eta\) and \(h_{\theta^+}(k)\ge c_b\). \end{enumerate} The target and pair may depend on \(M,K,\eta\). They are fixed before any experimental repetitions. In particular, the worst expected absolute prediction error over this pair is at least \((c_b-\eta)/2\), even when the common moments and eventual tail parameters are supplied exactly. For the following results put \[ u_0=\frac{1+\sqrt{1-8b}}2,\qquad a_0=\frac b{u_0},\qquad q=4a_0<1. \tag{2.8} \] \textbf{Theorem 2.2 (uniform-tail characterization).} For any nonempty class \(\mathcal F\) of laws satisfying (2.1) with the same \(m_0,b\), put \[ e_{\mathcal F}(N)=\sup_{\theta\in\mathcal F}A_\theta(N),\quad F_{\mathcal F}(T)=\sup_{\theta\in\mathcal F,\ C\text{ finite}}\Delta_\theta^C(T),\quad F_{\mathcal F}^{\rm site}(T)=\sup_{\theta\in\mathcal F}\Delta_\theta^{\{0\}}(T). \] Then the following conditions are equivalent: \[ \begin{split} e_{\mathcal F}(N)\longrightarrow0 &\quad\Longleftrightarrow\quad \sup_{\theta\in\mathcal F}\sum_{\ell\ge N}w_\ell\theta_\ell\longrightarrow0\\ &\quad\Longleftrightarrow\quad F_{\mathcal F}^{\rm site}(T)\longrightarrow0 \quad\Longleftrightarrow\quad F_{\mathcal F}(T)\longrightarrow0. \end{split} \tag{2.9} \] Section 4.2 proves the equivalence. An evaluable common envelope turns this qualitative criterion into a numerical history cap. \textbf{Theorem 2.3 (prediction intervals under a common envelope).} Under (2.1) and (2.6), there is an explicit function \(V_E(N)\le a_0\), tending to zero on positive integers, for which \(p_\theta(C)\le V_E(|C|)\) for every nonempty finite \(C\). For a nonempty finite target \(C\), if a supplied law \(Q\) on complete damage words of \(W=[\min C-T,\max C]\) satisfies \(\|P_{\theta,W}-Q\|_{\rm TV}\le\delta\) for every compatible \(\theta\), define, using \(T\) independent \(Q\) words, \[ z_Q=\Pr_Q(U_T=\varnothing),\qquad r_Q=\mathbb E_Q[\mathbf1_{\{U_T\ne\varnothing\}}V_E(|U_T|)], \qquad e_T=\min(1,T\delta). \] Then every compatible law satisfies \[ p_\theta(C)\in [\max(0,z_Q-e_T),\ \min(1,z_Q+r_Q+e_T)]. \tag{2.10} \] Total variation throughout is the supremum of event-probability differences. Under (2.7), one may use \[ V(N)=\min(a_0,C_pN^{-\epsilon}),\qquad C_p=c_p(b)B,\quad c_p(b)=\frac{q(1+q)}{b(1-q)^3}. \tag{2.11} \] \textbf{Theorem 2.4 (sharp class-uniform physical-history scale).} For fixed \(b\in(0,1/32]\), there are explicit finite constants \(0M\), those origin intervals overlap and \[ U_\ell(A)=\ell+\operatorname{diam}(A). \tag{3.2} \] Consequently a complete diameter-\(M\) word depends on the long intensities only through \(\sum_{\ell>M}\theta_\ell\) and \(\sum_{\ell>M}\ell\theta_\ell\), once the intensities at lengths at most \(M\) are specified. Void probabilities for all subsets determine every binary pattern by inclusion--exclusion. This is the lattice counterpart of the Boolean capacity calculation in {[}LP18, Theorem 17.3{]}. Matching the second length moment adds information without changing this local equality. There is no conflict with identification from all lags. Set \(\lambda=\sum_\ell\ell\theta_\ell\), \(\Gamma_d=\sum_{\ell>d}(\ell-d)\theta_\ell\), and \(Y_x=\mathbf1_{\{x\in D\}}\). Directly from (3.1), \[ \operatorname{Cov}(Y_0,Y_d)=e^{-2\lambda}(e^{\Gamma_d}-1), \qquad \theta_\ell=\Gamma_{\ell-1}-2\Gamma_\ell+\Gamma_{\ell+1}. \tag{3.3} \] Thus exact covariances at every lag identify the intensity sequence inside this model class. The covariance identity is the lattice counterpart of {[}LP18, Theorem 17.8{]}. Theorem 17.10 there identifies a continuum Boolean model's intensity and centred grain law from its full capacity functional; here the second-difference formula proves identification within the declared lattice-interval class. The obstruction concerns a bounded window. \textbf{Proposition 3.1 (equality of adaptive local transcripts).} Suppose two laws have the same complete damage-word distribution on every window of diameter at most \(M\). Every \(M\)-local observation policy defined in Section 2.2 has the same transcript law under the two models. The conclusion includes all measurable forecasts based on the transcript and the policy's independent random seed. \textbf{Proof.} Include the random seed in the transcript and couple it identically. Conditional on any matched prefix, the next decision rule chooses the same window and continuation or stopping decision in both models. If it requests a fresh row, the conditional word law is the same supplied marginal, since the row is independent of the prefix. Couple that word identically. Every recovery readout and every arm using it is then the same measurable function of the matched history. This inductively couples every finite prefix identically, including a stopped transcript extended by a fixed terminal symbol. Equality of all finite-prefix laws gives equality of countable transcript laws, hence of any measurable output. Reusing a row within its already supplied window is postprocessing and introduces no new kernel. Repeating experiments on fresh rows repeats this argument. \(\square\) \subsubsection{3.2. A three-moment construction}\label{a-three-moment-construction} Let \(D=M+1\), \(K_D=(5D^2+2D)/2\), \(s=b/(2K_D)\), and \[ S_p=\sum_{\ell\ge m_0}w_\ell\ell^{-p},\qquad a=b/(2S_p). \] Take \(q^-=(\delta_D+\delta_{3D})/2\). For an integer \(L\ge\max(2k,3D)\), put \(d=L-2D\) and \[ q^+=\frac D{d+D}\delta_D+\left(1-\frac Dd\right)\delta_{2D} +\frac{D^2}{d(d+D)}\delta_L. \tag{3.4} \] Both are probability distributions with first moment \(2D\) and second moment \(5D^2\). For \(q^+\), the centred first moment about \(2D\) is zero and the centred second moment is \(D^2\): the two noncentral weights satisfy \(w_D=(d/D)w_L\). Set \[ \theta_\ell^\pm=a\ell^{-p}+s q^\pm(\ell). \tag{3.5} \] All added atoms lie above \(M\). The zeroth, first and second moments match, and \(\chi=aS_p+sK_D=b\). Equations (3.1)--(3.2) give exact complete-word equality. After their finitely many additional lengths, both laws are exactly \(a\ell^{-p}\). An interval \([z,z+L-1]\) born \(j\) rows earlier leaves, under erosion alone, the remnant \([z+j,z+L-1]\). Since (2.2) is monotone, this remnant remains contained in the actual holes, irrespective of other damage. For a target \([0,k-1]\), counting origins and ages gives \[ h_\theta(k)\ge1-\exp\left[-\theta_L \frac{(L-k+1)(L-k+2)}2\right]. \tag{3.6} \] For the added length-\(L\) atom of \(\theta^+\), \(L\ge2k\) and \(d(d+D)\le L^2\) make the exponent at least \(sD^2/8\ge b/48\). Hence \(h_{\theta^+}(k)\ge c_b\). The low law \(\theta^-\) is fixed before choosing \(k\). Section 4 proves that its stationary block probabilities tend to zero whenever \(\chi<\infty\) and the small-budget bound holds. Choose \(k\ge K\) with \(h_{\theta^-}(k)\le\eta\), then choose \(L\). This proves Theorem 2.1 without a stationary amplitude theorem or a regular-variation limit. If a randomized forecast \(Z\) has the same output law under two targets \(h_-0\). Choose an integer \(K_0\ge1\) with \(V_E(K_0+1)\le\tau/3\), put \(\gamma=\tau/[3a_0(K_0+1)^2]\), and choose \[ N_0\ge\lceil(m_0+1)/2\rceil,\quad E_{N_0}\le\gamma/8, \quad R_0=2N_0-1,\quad \beta_0=\frac{\gamma}{2\sum_{\ell=m_0}^{R_0}\ell},\quad \eta_*=\left(\frac{\beta_0}{1+\beta_0}\right)^{K_0}>0. \tag{4.6} \] These choices exist because the envelope tends to zero. For every finite input set, the expected average occupation of sizes \(1,\ldots,K_0\) over \(s=0,\ldots,R-1\) is bounded by \[ \max\left\{\frac{K_0+1}{R}+\gamma(K_0+1)^2, \frac1{R\eta_*}\right\}. \tag{4.7} \] Here is a direct proof uniform in the law. Let \(R_D(C)=|C\cap D|\) count hit sites in a fresh row. For \(j=|C|\), \(\mathbb ER_D(C)\le j\lambda_\theta\), regardless of dependence among the hit indicators. If \(\lambda_\theta\le\gamma\), use the bounded potential \(F(C)=\min(|C|,K_0+1)\). For \(1\le j\le K_0\), no hit has probability at least \(1-j\lambda_\theta\) and raises the potential by at least one; on a hit its loss is at most \(j\). Thus its expected drift is at least \(1-j(j+1)\lambda_\theta\ge1-\lambda_\theta(K_0+1)^2\). For \(j\ge K_0+1\), a potential decrease requires at least \(j-K_0\) hit sites, and its magnitude is at most \(K_0+1\). Markov's inequality therefore gives \[ \begin{split} \mathbb E\big[(F(C)-F(G_D(C)))_+\big] &\le(K_0+1)\Pr(R_D(C)\ge j-K_0)\\ &\le(K_0+1)\frac{j\lambda_\theta}{j-K_0} \le\lambda_\theta(K_0+1)^2. \end{split} \] The empty state has zero drift. Combining the cases yields drift at least \(\mathbf1_{\{1\le|C|\le K_0\}}-\lambda_\theta(K_0+1)^2\). Telescoping and \(0\le F\le K_0+1\) give \[ \frac1R\sum_{s=0}^{R-1}\Pr_C(1\le|U_s|\le K_0) \le\frac{K_0+1}{R}+\lambda_\theta(K_0+1)^2, \] which proves the first term in (4.7). If \(\lambda_\theta\ge\gamma\), the inequality \[ \sum_{\ell\ge2N_0}w_\ell\theta_\ell\le4A_\theta(N_0) \le4E_{N_0}\le\gamma/2 \tag{4.8} \] follows from \((\ell-N_0+1)(\ell-N_0+2)\ge\ell(\ell+1)/4\). Since \(w_\ell\ge\ell\), some supported length at most \(R_0\) has intensity at least \(\beta_0\). At every visit to a nonempty set of at most \(K_0\) sites, require that length at each site-origin. These independent atoms kill the set in one row with probability at least \((1-e^{-\beta_0})^{K_0}\ge\eta_*\). Absorption events at distinct times are disjoint, so their total probabilities sum to at most one. This proves the second term in (4.7). The absorption formula makes the true remainder nonincreasing in time. For each law and seed, (4.4) gives \[ \Delta_\theta^C(R)\le V_E(K_0+1)+\frac{a_0}{R} \sum_{s=0}^{R-1}\Pr_C(1\le|U_s|\le K_0). \] Using (4.7) therefore gives \(\sup_{\theta,C}\Delta_\theta^C(R)\le\tau\) whenever \[ R\ge\max\{1,3a_0(K_0+1)/\tau,3a_0/(\tau\eta_*)\}. \tag{4.9} \] The zero law belongs to the low-intensity branch. This proof establishes a last-cut guarantee over all finite targets without assuming a monotone realized ancestor size or a monotone expected cardinality certificate. Evaluating the cap requires numerical access to the supplied envelope and conservative arithmetic. Section 6 derives the polynomial rate under the moment restriction without using this cap. \subsubsection{4.2. The exact uniform-tail criterion}\label{the-exact-uniform-tail-criterion} We prove (2.9). Set \(Q_{\mathcal F}(N)=\sup_{\theta\in\mathcal F}\sum_{\ell\ge N}w_\ell\theta_\ell\). The coefficient bound used in (4.8) and the definition of \(A_\theta\) give \[ Q_{\mathcal F}(2N)\le4e_{\mathcal F}(N),\qquad e_{\mathcal F}(N)\le Q_{\mathcal F}(N). \tag{4.10} \] Hence either uniform tail condition implies the other. If \(e_{\mathcal F}(N)\to0\), take \(E_1=b\) and \(E_N=e_{\mathcal F}(N)\) for \(N\ge2\); Section 4.1 gives \(F_{\mathcal F}(T)\to0\). This implies the corresponding single-site convergence. Conversely, length-\(\ell\) atoms at ages at least \(T\) whose individual erosion remnant covers zero have total intensity \(A_\theta(T+1)\). These old atoms are independent of the latest \(T\) rows, and their remnants are contained in the actual hole field. Thus \[ \Delta_\theta^{\{0\}}(T) \ge(1-p_{\theta,T}(\{0\}))(1-e^{-A_\theta(T+1)}) \ge\frac{1-a_0}{1+b}A_\theta(T+1). \tag{4.11} \] Here \(p_{\theta,T}(\{0\})\le a_0\), \(A_\theta(T+1)\le b\), and \(1-e^{-x}\ge x/(1+b)\) for \(0\le x\le b\). Taking the supremum gives \[ F_{\mathcal F}^{\rm site}(T)\ge\frac{1-a_0}{1+b}e_{\mathcal F}(T+1). \] Thus single-site uniform forgetting forces \(e_{\mathcal F}(N)\to0\), closing all implications. No maximizing law is required. This proof concerns a tail restriction on the law class itself; it does not assume that bounded calibration identifies that restriction. \subsection{5. A boundary comparison valid for the entire moment class}\label{a-boundary-comparison-valid-for-the-entire-moment-class} \subsubsection{5.1. Flux and a comparison distribution}\label{flux-and-a-comparison-distribution} Write \(h_k=h_\theta(k)\), \(u=u_\theta\), and \[ d_k=h_k-h_{k+1} =\Pr([-k+1,0]\subset H_\theta,\ 1\notin H_\theta). \tag{5.1} \] The \(d_k\) are nonnegative and nonincreasing, and \(\sum_{k\ge1}d_k=h_1\). Define a nonnegative integer variable \(Z_\theta\) by \[ \Pr(Z_\theta\ge k)=d_k/u,\quad k\ge1. \tag{5.2} \] These are valid tails: \(d_1/u\le a_0/u_0<1\) and \(d_k\to0\). In particular, \[ \mathbb EZ_\theta=h_1/u\le\mu_b:=b/u_0^2<1/8. \tag{5.3} \] For \(k\ge2\), monotonicity gives \(\lfloor k/2\rfloor d_k\le h_{\lceil k/2\rceil}\). Therefore, under (2.7), \[ \Pr(Z_\theta\ge k)\le L k^{-1-\epsilon},\qquad L=3\,2^\epsilon C_p/u_0. \tag{5.4} \] The bound also holds at \(k=1\), since \(C_p\ge4B/u_0\) and \(L\ge12\mu_b\). Thus \(\Pr(Z_\theta>y)\le L y^{-1-\epsilon}\) for all real \(y>0\). \subsubsection{5.2. Conditional boundary kernels and single-time domination}\label{conditional-boundary-kernels-and-single-time-domination} Insert the right hole ray \([0,\infty)\) into the stationary baseline. Let \(\Lambda_t\) be the left endpoint of its right-infinite hole component after \(t\) fresh rows. The ray-constrained open paths are exactly those whose coordinate at row zero is at most \(-1\). Thus the deterministic checkpoint in Lemma B.1 is row zero with cap \(c=-1\). At row \(t\) every site in \([t,\infty)\) is deterministically unreachable. Put \(X_t=t-\Lambda_t\) and \(J_t=\Lambda_t+1-\Lambda_{t+1}\ge0\). \textbf{Corollary 5.1 (ray specialization of the checkpoint comparison).} At every fixed time \(t\ge0\) and every boundary value \(x\) of positive probability, \[ \Pr(J_t\ge k\mid\Lambda_t=x)\le d_k/u,\qquad k\ge1. \tag{5.5} \] \textbf{Proof.} To display the prehistory explicitly, start fully open at row \(-n\). Write \(H_t^{0,n}\) for the unrestricted holes and \(H_t^{c,n}\) for the holes after requiring coordinate at most \(-1\) at row zero. This latter restriction inserts the ray into the finite-prehistory baseline at that row. For \(z=x-1\) define \[ A_z^n=\{z\notin H_t^{0,n}\},\quad O_z^n=\{z\notin H_t^{c,n}\},\quad R_x^n=\{[x,t-1]\subset H_t^{c,n}\}. \] For \(x\le t\), the event \(\{\Lambda_t^n=x\}\) is precisely \(O_z^n\cap R_x^n\); \(R_t^n\) is certain. Both events read finitely many terminal coordinates. On \(O_z^n\), Lemma B.1(i) identifies the old fields strictly left of \(z\). On \(O_z^n\cap R_x^n\), the eroded constrained ray starts at \(x+1\), while \(z\) and \(x\) are open in the eroded field. For any hole field \(H\), let \[ f_{k,x}(H)=\Pr_D\big([x-k+1,x]\subset(E(H)\setminus\{z,x\})\cup D\big), \] where \(D\) is an independent fresh row. Clamping \(z,x\) open makes \(f_{k,x}\) a bounded increasing function only of old hole coordinates strictly left of \(z\): the other queried eroded sites are at most \(z-1\), and erosion there reads that site and its left neighbour. On the boundary event, this clamp changes nothing in the constrained eroded exterior. Consequently \[ \Pr(J_t^n\ge k\mid\Lambda_t^n=x) =\mathbb E[f_{k,x}(H_t^{0,n})\mid O_z^n,R_x^n]. \] Part (ii) of Lemma B.1 first removes the right increasing hole event \(R_x^n\). Part (iii) then replaces constrained openness by unrestricted openness. They give, in this order, \[ \mathbb E[f_{k,x}(H_t^{0,n})\mid O_z^n,R_x^n] \le\mathbb E[f_{k,x}(H_t^{0,n})\mid O_z^n] \le\mathbb E[f_{k,x}(H_t^{0,n})\mid A_z^n]. \tag{5.6} \] Now let \(n\to\infty\) under the common increasing-prehistory construction. For fixed \(t,x,k\), all fields and observables just used depend on finitely many terminal sites. Their indicators and expectations converge. If \(\Pr(\Lambda_t=x)>0\), the probabilities of \(O_z^n\cap R_x^n\) converge to a positive number and are eventually positive, as are those of \(O_z^n\); \(\Pr(A_z^n)\to u>0\). Passing the conditional ratios to the limit is therefore valid. The unrestricted terminal field has the stationary law of \(H_\theta\). Finiteness of the ray boundary follows as in Appendix B.2: sufficiently far to the permitted side it agrees with a positive-density stationary open field, while the other side contains a deterministic unreachable ray. Put \(B=E(H_\theta)\) and \(\Gamma=\{z,x\notin B\}\). The event \(A_z=\{z\notin H_\theta\}\) is contained in \(\Gamma\). On \(\Gamma\) clamping is immaterial, so the last conditional expectation in the stationary limit of (5.6) is at most \[ \frac{\Pr(\Gamma,\ [x-k+1,x]\subset B\cup D)}u \le\frac{h_k-h_{k+1}}u. \tag{5.7} \] Indeed \(B\subset B\cup D\), whose law is \(H_\theta\), and \(B(x)=0\) on \(\Gamma\) excludes the event that the block was already contained in \(B\). The probabilities of the new and old block events are \(h_k\) and \(h_{k+1}\). This proves (5.5). \(\square\) We transfer the conditional increment bounds to fixed-time positions through an auxiliary chain with the same one-time marginals. Define \[ K_t(x,\cdot)=\mathcal L(J_t\mid X_t=x) \] at states of positive marginal probability, and set \(K_t(x,\{0\})=1\) at the other nonnegative integer states. On this countable state space the kernel and its quantile are measurable. For \(00\) and \(\delta_0\ge1/16\). The threshold \(n_0\) can grow rapidly as \(\epsilon\) decreases; the final multiplier is controlled through \(n_0^\epsilon\), as shown in Section 6.3. Truncate the independent comparison variables at \(rT\): \(Y_i=Z_i\mathbf1_{\{Z_i\le rT\}}\). The event of an omitted large summand costs at most \(Lr^{-1-\epsilon}T^{-\epsilon}\). Since \(Y_i^2\le rTY_i\) and \(\mathbb EY_i\le\mu_b\), Chebyshev yields \[ \Pr\left(\sum_{i=1}^TZ_i>(\mu_b+d_1)T\right) \le Lr^{-1-\epsilon}T^{-\epsilon}+\frac{\mu_br}{d_1^2} \le Lr^{-1-\epsilon}T^{-\epsilon}+2\delta_0/3. \tag{6.2} \] Splitting \(X_0^\circ\) at \(d_0T\) and using (5.8) bounds either half-line's retreat beyond \(T/8\) by \[ \Phi(T)=J_\epsilon T^{-\epsilon}+2\delta_0/3. \tag{6.3} \] Only a fixed small failure probability is needed. At \(T=8n-1\) and \(n\ge n_0\), \(\Phi(T)\le\delta_0\). No endpoint logarithm or moment above order \(2+\epsilon\) enters this step. \subsubsection{6.2. One selected point and one output interval}\label{one-selected-point-and-one-output-interval} For ordinary propagation \(F_D(C)=(C\cup(C-1))\setminus D\), the following deterministic path statement supplies the geometric input to the restart. \textbf{Lemma 6.1 (prefix exchange for a surviving source).} Fix an integer \(T\ge0\) and any common disaster rows \(D_1,\ldots,D_T\). Let \(A_T^S\) be the sites reachable from a source set \(S\) at row zero by open paths with increments \(0,-1\). If \(A_T^{\{y\}}\ne\varnothing\), write \(L=\min A_T^{\{y\}}\) and \(R=\max A_T^{\{y\}}\). Then \[ L=\min A_T^{[y,\infty)},\qquad R=\max A_T^{(-\infty,y]},\qquad A_T^{\{y\}}=A_T^{\mathbb Z}\cap[L,R]. \tag{6.3a} \] \textbf{Proof.} Two allowed paths whose endpoint order reverses their starting order meet at an integer vertex: the difference of their coordinates changes by at most one at each step. At a meeting vertex, the prefix of either open path can be concatenated with the suffix of the other. If a path from \([y,\infty)\) ended strictly left of \(L\), it would meet a path from \(y\) to \(L\). Swapping prefixes would give a path from \(y\) to that smaller endpoint, a contradiction. Inclusion of the point source in the half-line gives the reverse inequality and proves the first equality. A path from \((-\infty,y]\) ending strictly right of \(R\) meets a path from \(y\) to \(R\) and gives the symmetric contradiction, proving the second equality. Now take \(w\in A_T^{\mathbb Z}\cap[L,R]\) and an open path to \(w\) starting at \(q\). If \(q\le y\), compare it with a path from \(y\) to \(L\): their starting and ending orders force a meeting. If \(q\ge y\), compare it with a path from \(y\) to \(R\). In either case the prefix from \(y\) and the suffix to \(w\) form an open path, so \(w\in A_T^{\{y\}}\). The reverse inclusion is immediate. All statements are pathwise and require only nonemptiness at the declared finite time. \(\square\) For an interval \(J\) of \(n\) sites, the full-open finite-history baseline has expected hole count at most \(na_0\). Markov's inequality gives \[ \Pr(|O^{\rm full}\cap J|<\lceil n/2\rceil)\le2a_0. \tag{6.4} \] There is no conditioning on survival in (6.4). For actual ancestors, the exact order over \(8n\) rows is \[ G_{D_{8n}}\cdots G_{D_1}(\{y\}) =B_- F_{D_{8n}}\cdots F_{D_2}(\{y\}\setminus D_1), \qquad B_-A=A\cup(A-1). \tag{6.5} \] If the first kill leaves \(y\) alive, the \(8n-1\) ordinary rows are fresh and independent of that kill. Choose \(a\) with \(y\in[a,a+n-1]\) and let \(J=[a-3n,a-2n-1]\). On both half-line events with retreat at most \((8n-1)/8\), the ordinary left endpoint is at most \[ y-7(8n-1)/8a-n\). They enclose \(J\). The final expansion only adds sites. Therefore \[ \Pr_y(0<|U_{8n}|<\lceil n/2\rceil) \le2\Phi(8n-1)+2a_0. \tag{6.6} \] For an arbitrary nonempty finite \(C\), choose one point \(y(C)\) before revealing the next rows. The update preserves unions. If \(C\) has a small nonempty descendant, the chosen point either died or also has a small nonempty descendant. Its extinction probability by that time is at most its eventual extinction probability \(h_\theta(1)\le a_0\). Hence, uniformly in \(C\) and its geometry, \[ \Pr_C(0<|U_{8n}|<\lceil n/2\rceil) \le3a_0+2\Phi(8n-1)\le\eta_0,\qquad n\ge n_0. \tag{6.7} \] The selected family may die while another family survives; the extra \(a_0\) pays exactly that event. Conditional on any predictable seed, the future rows still have their original independent law. \subsubsection{6.3. Remaining-time induction}\label{remaining-time-induction} Let \(F(T)=\sup_{\theta\in\mathcal K,C\text{ finite}}\Delta_\theta^C(T)\). Apply (4.5), split at \(|U_{8n}|=\lceil n/2\rceil\), use compression on the large branch and (6.7) on the small nonempty branch. For \(t\ge1,n\ge n_0\), \[ F(t+8n)\le C_p\lceil n/2\rceil^{-\epsilon}+\eta_0F(t). \tag{6.8} \] Put \[ L_0=16(n_0+1),\qquad K=\max\{2C_p64^\epsilon,\ a_0L_0^\epsilon\}. \tag{6.9} \] For an integer \(Q\ge L_0\), take \(n=\lfloor Q/16\rfloor\) and \(t=Q-8n\). Then \(n\ge n_0\), \(\lceil n/2\rceil\ge Q/64\), and \(Q/2\le t0\) such that for every integer \(D\ge D_0\), two laws in \(\mathcal K\) have identical diameter-\((D-1)\) calibration contracts, equal primitive moments of orders zero, one and two, and the same eventual power-law parameters, but at \(k_D=\lceil7D/2\rceil\) satisfy \[ h_+(k_D)-h_-(k_D)\ge\frac{\zeta}{16(1+b)}D^{-\epsilon}. \tag{7.1} \] \textbf{Proof.} Use the background \(\theta^0\) in Section 6.4, put \(s_D=\zeta D^{-2-\epsilon}\) and take \[ \theta^- =\theta^0+s_D\delta_D+3s_D\delta_{3D},\qquad \theta^+ =\theta^0+3s_D\delta_{2D}+s_D\delta_{4D}. \tag{7.2} \] Both add-ons have mass \(4s_D\), first moment \(10s_DD\), second moment \(28s_DD^2\) and recovery budget \(c_D=s_D(14D^2+5D)\le19\zeta D^{-\epsilon}\). Their higher-moment coefficients are at most \(88\zeta\le B/2\), and \(c_D\le b/4\). For \(D\ge m_0\), the laws belong to \(\mathcal K\). Equations (3.1)--(3.2) prove the calibration equality. The high law's length-\(4D\) atom gives \(A_+(k_D)\ge s_DD^2/8\), and hence \[ h_+(k_D)\ge\zeta D^{-\epsilon}/[8(1+b)]. \tag{7.3} \] For the low law, dominate its tree by one with activity exactly \(a_0\). Conditional on a positive node, give it a background label with probability \(\chi_0/b\) and background cost tail \(A_0(N)/\chi_0\); an add-on label with probability \(c_D/b\) and cost tail \(A_{\rm add}(N)/c_D\), supported on \([1,3D]\); or a dummy label with the remaining probability and cost one. For \(N\ge2\) its joint active-cost tail is \(A_-(N)/u_0\ge A_-(N)/u_-\); at \(N=1\) activity \(a_0\) dominates \(\chi_-/u_-\). Thus Lemma 4.1's node coupling applies. Let \(J\) be its number of positive nodes, \(\mu=\mathbb EJ=a_0/(1-2a_0)\), and \[ s_2=\mathbb EJ^2=\frac{a_0(1+4\mu+2\mu^2)}{1-2a_0},\qquad f_2=s_2-\mu. \] Put \(V_i=0\) at an add-on label and equal to its cost otherwise. Tail summation, or \(W_0\le\ell\) in the length-biased background mixture, gives \(\chi_0\mathbb EW_0^2\le\sum_\ell\theta^0_\ell w_\ell\ell^2\le H_0\). Hence \(\mathbb EV_i^2\le m_2=1+H_0/b\). If total cost is at least \(k_D\) and at most one add-on node occurs, the non-add-on sum is at least \(D/2\). Conditional on \(J=j\), its second moment is at most \(j^2m_2\). Two or more add-on labels cost at most \(f_2(c_D/b)^2/2\). Therefore \[ h_-(k_D)\le C_{\rm bg}D^{-2}+C_{\rm pair}\zeta^2D^{-2\epsilon}, \quad C_{\rm bg}=4m_2s_2,\quad C_{\rm pair}=\tfrac12f_2(19/b)^2. \tag{7.4} \] Take \(D_0\) to be the ceiling of the maximum of \(m_0,1\) and \[ [32(1+b)C_{\rm bg}/\zeta]^{1/(2-\epsilon)},\qquad [32(1+b)C_{\rm pair}\zeta]^{1/\epsilon}. \] Each term in (7.4) is then at most \(\zeta D^{-\epsilon}/[32(1+b)]\). Subtract from (7.3) to prove (7.1). \(\square\) Equation (3.7) gives a worst expected absolute error of at least half of (7.1), even with the exact local word law. If an interval has coverage at least \(1-\alpha\) under each law, \(\alpha<1/2\), its common output distribution satisfies \[ \Pr(|I|\ge h_+(k_D)-h_-(k_D))\ge1-2\alpha. \tag{7.5} \] This is a one-sided spatial extrapolation bound at a declared growing target. The common moment budget makes its gap decay, while Theorem 2.1 retains a constant gap over a broader tail class. \subsection{8. Prediction procedures and distinct resource accounts}\label{prediction-procedures-and-distinct-resource-accounts} \subsubsection{8.1. Complete-word calibration}\label{complete-word-calibration} For a nonempty finite target \(C\) under (2.7), choose \(T\) with \(KT^{-\epsilon}\le\tau\), using (6.9). If \(Q\) approximates the true complete row word on \([\min C-T,\max C]\) within \(\delta\), uniformly over compatible laws, then \[ p_\theta(C)\in[\max(0,z_Q-e_T),\ \min(1,z_Q+\tau+e_T)]. \tag{8.1} \] Row coupling transfers the emptiness probability, and the true stationary remainder lies in \([0,\tau]\). The bound covers every finite target whose causal window is supplied. Its accuracy input is total variation of the complete joint word on that window; low-order moments or separate site marginals do not determine this input. A corresponding result applies to a shared-history preparation. At row zero delete disjoint finite root sets \(I_1,I_3\) from the stationary open field. Let \(X_A\) be the terminal all-hole indicator for a finite target \(C\) after deleting roots indexed by \(A\subset\{1,3\}\) and evolving for age \(t\). The four-arm contrast \[ D=X_{13}-X_1-X_3+X_\varnothing\in\{0,1\} \tag{8.2} \] uses the same old field and future rows. An empty target has contrast zero and needs no prediction. For a nonempty target, reverse the \(t\) future rows first, remove the roots, then reverse the same \(T\) older rows for every arm. The cut sets satisfy \(C_{13}\subset C_1\cap C_3\) and \(C_\varnothing=C_1\cup C_3\). With an all-open cut field the contrast is \[ d=\mathbf1_{\{C_{13}=\varnothing,C_1\ne\varnothing,C_3\ne\varnothing\}}. \] For any recent history, discrepancy from the true contrast has conditional probability at most \[ \pi(C_{13})+\pi(C_1)+\pi(C_3),\qquad \pi(S)=\mathbf1_{\{S\ne\varnothing\}}p_\theta(S). \tag{8.3} \] If \(C_{13}\) is nonempty, a positive true response requires it all holes. If it is empty, disagreement requires one of the two nonempty single-arm sources to be all holes. Equality of a single-arm source with \(C_{13}\) gives exact response zero. These cases prove (8.3). Condition on future rows and preparation. Older rows are independent of the resulting seeds, so each of the three expectations in (8.3) is a true remainder at time \(T\), at most \(\tau\). They need not be independent of one another. If the complete word on \([\min C-(t+T),\max C]\) is within \(\delta\) of \(Q\), and \(d_Q=\mathbb E_Qd\), then \[ \mathbb E_\theta D\in [\max(0,d_Q-3\tau-e_{t+T}),\ \min(1,d_Q+3\tau+e_{t+T})]. \tag{8.4} \] \subsubsection{8.2. Direct fresh-row observations}\label{direct-fresh-row-observations} The inputs to this protocol are the declared class parameters and observed joint row words; replay uses the known deterministic update. In contrast, the kernel-based constructions in {[}DLL25, Section 2{]} specify lower-envelope transition probabilities over admissible pasts. Fixed-diameter calibration does not supply those kernel inputs. One can avoid estimating an entire word distribution by observing fresh words and replaying them. Fix in advance \(J=[a,b_0]\) of \(d\) sites and a maximum age \(t_{\max}\). Consider all hole targets \(C\subset J\) and all response queries with \(C\subset J\), disjoint \(I_1,I_3\subset J\) and \(0\le t\le t_{\max}\). There are at most \[ m=2^d+(t_{\max}+1)6^d\le(t_{\max}+2)6^d \tag{8.5} \] indexed queries. The \(6^d\) counts two target memberships and three root labels per site. For \(0<\eta,\alpha<1\), put \(\rho=\eta/6\), \(s=\eta/2\), \[ \begin{gathered} T=\max(1,\lceil(K/\rho)^{1/\epsilon}\rceil),\quad H=T+t_{\max},\quad W=[a-H,b_0],\\ N=\left\lceil2\eta^{-2}\left[d\log6+ \log\frac{2(t_{\max}+2)}\alpha\right]\right\rceil. \end{gathered} \tag{8.6} \] Read \(NH\) independent fresh row words on \(W\) and divide them into \(N\) disjoint chronological blocks. Each block's first \(T\) rows supplies old history and its remaining rows supply future history. Reverse, delete roots and reverse the old rows in the order specified above. All queries may reuse a block; distinct blocks use disjoint rows. Hole outputs are Bernoulli emptiness indicators. Response outputs are also Bernoulli. To see this directly, split the row-zero ancestor set into \(A_1=S\cap I_1\), \(A_3=S\cap I_3\) and \(E=S\setminus(I_1\cup I_3)\). For a fixed finite-history hole field put \(Q_Z=\mathbf1_{\{Z\subset H_T\}}\). The response is \[ Q_E(1-Q_{A_1})(1-Q_{A_3})\in\{0,1\}. \tag{8.7} \] The hole bias lies in \([0,\rho]\); response absolute bias is at most \(3\rho\). For a Bernoulli variable \(Y\), the log moment-generating function of \(Y-\mathbb EY\) has second derivative at most \(1/4\), so it is bounded by \(\lambda^2/8\). Chernoff and independent blocks give \(\Pr(|\bar Y-\mathbb EY|>s)\le2e^{-2Ns^2}\). A union bound over (8.5) gives one probability-\(1-\alpha\) event on which every hole interval \[ [\max(0,\bar Y-s),\ \min(1,\bar Y+s+\rho)] \tag{8.8} \] and every response interval \[ [\max(0,\bar Y-s-3\rho),\ \min(1,\bar Y+s+3\rho)] \tag{8.9} \] covers its target. The statement holds under whichever admissible law generates the observations. Queries may be chosen after seeing the data within the declared domain. Enlarging that domain or stopping early requires additional confidence control. The procedure reads \(NH\) row words and \(NH(d+H)\) site values. A one-step all-open experiment realizes each damage-word observation, provided \(W\) and its left predecessor are prepared open. Such an implementation uses up to \(NH\) resets. Replaying one requested query requires at most order \(4NH(d+H)\) site operations. At fixed class parameters, domain, age cap and confidence, the sufficient row and site counts are respectively \(O(\eta^{-2-1/\epsilon})\) and \(O(\eta^{-2-2/\epsilon})\). These are sufficient costs for the stated fresh-row protocol. The matching optimization in Theorem 2.4 concerns the deterministic all-open history depth. \subsubsection{8.3. Continuity for all-hole events}\label{continuity-for-all-hole-events} Let \(\theta,\theta'\) belong to the same fixed class \(\mathcal K\), and set \(d_w(\theta,\theta')=\sum_\ell w_\ell|\theta_\ell-\theta'_\ell|\). Couple common Poisson atoms and independent difference atoms. At an explored set \(U_s\), difference atoms hitting it have mean at most \(\frac23|U_s|d_w\), because \(\ell/w_\ell\le2/3\). Before the first discrepancy \(|U_s|\le|C|(s+1)\). Thus the probability of a first discrepancy by row \(T\) is at most \(|C|T(T+1)d_w/3\), regardless of target diameter. For \(|C|\le T\), the two remainders lie in \([0,KT^{-\epsilon}]\), so their difference costs at most \(KT^{-\epsilon}\). For \(|C|>T\), both stationary probabilities are at most \(C_pT^{-\epsilon}\). Consequently, with \(K_m=\max(K,C_p)\), \[ \sup_{C\text{ finite}}|p_\theta(C)-p_{\theta'}(C)| \le\min\{1,K_mT^{-\epsilon}+(2/3)T^3d_w\}. \tag{8.10} \] For \(0\gamma_t\), a covering interval either intersects \(\gamma_t\), in which case it is absent in \(\omega\), or has its entire interval strictly right of \(\gamma_t\). In the latter case it is selected in both configurations and receives the same entry \(y_e\); openness of \(p\) forces that entry to be 1. Connectedness of the interval exhausts these cases. Thus \(q\) is open, proving the order of the new rightmost paths. The leftmost paths are unchanged. Types strictly left of \(\alpha\) are unchanged in both configurations. A type strictly right of the new rightmost path in \(\omega\) is also strictly right of \(\gamma\) and \(\alpha\), so it was selected in both configurations and now has equal values. These observations verify all three comparison conditions after the right update. For a common left update, set \(\alpha=\widehat R\le R=\gamma\). Given a path \(p\) open after the update in \(\omega\), check \(q=\min(p,\alpha)\) in the updated \(\widehat\omega\). At a site of \(\alpha\), all covering types are absent. At a site \(p_t<\alpha_t\), a covering interval either meets \(\alpha_t\) and is absent in \(\widehat\omega\), or is strictly left of both \(\alpha_t\) and \(\gamma_t\) and receives the same entry \(y_e\). This proves that \(q\) is open and orders the new leftmost paths. The rightmost paths are unchanged. Every type strictly left of the new \(\widehat L\) is strictly left of both old rightmost paths, so it was identically resampled. Types strictly right of \(\gamma\) are unchanged in both configurations. Hence this update also preserves all three conditions, including when the two selected type sets differ. For a single right update from identical input configurations, decreasing the resampled absence variables adds only right-side disasters. The old leftmost path stays open, and the new rightmost path can only move left; all types beyond the larger rightmost path have the required absence order, while types left of the common leftmost path are unchanged. This establishes the comparison after that update. For a single left update, increasing the resampled absences adds only left-side open paths; the rightmost path stays fixed and the leftmost path can only move left, with the required order on the two outer type sets. The common-update preservation just proved propagates either comparison through subsequent updates. Finally, if \(\widehat\omega\preceq\omega\), every path to a terminal \(x<0\) open in \(\omega\) yields an open path in \(\widehat\omega\) by taking its minimum with \(\widehat L\). An interval covering a strictly left point either intersects \(\widehat L\) or belongs to its strictly left type set, where condition (ii) applies. Terminal reachability on the left therefore increases, while reachability on the right decreases by the corresponding maximum construction and condition (iii). Realize the resampling chain on a product space by sampling arrays \(Y^R_{k,e}\) and \(Y^L_{k,e}\) in advance for every update index \(k\) and every retained interval type \(e\). Use an entry only when its type belongs to the currently selected resampling set; otherwise ignore it. All these entries are independent with the original type-specific absence probabilities. If one entry is changed, the two chains have identical histories before its update. At that update either the entry is ignored in both chains or the single-update comparison established above applies. Every subsequent common update preserves that comparison, even when the two resampling sets differ. Thus, after any fixed number of updates, every left increasing hole functional is decreasing in the left-update entries and increasing in the right-update entries. Right increasing functionals have the opposite signs. Complementing the right-update entries places the first functional and the negative of the second in the same coordinatewise decreasing class. The product-measure association inequality gives their nonnegative covariance, equivalently the desired negative correlation. A direct proof for the type-dependent parameters used here is given below. Start at the all-1 configuration and let the finite conditional chain converge to \(\mu\) to obtain the asserted conditional inequality. For the infinite model, first fix the time and the finite terminal coordinates of the functions. Their backward cones are finite. Remove types of length greater than \(K\); the total mean of omitted types intersecting the cones is at most a constant times \(\varrho T\sum_{r>K}\nu_r(r+C)\), which tends to zero by \(S_1<\infty\). The conditional event has probability at least that of an open vertical path, \(e^{-\varrho TS_1}>0\). Thus the conditional inequality passes to the limit. Deterministic absence variables can be removed or obtained by parameter limits. The same argument applies to the stated type-dependent parameters and last-row interpolation. \(\square\) The product-measure correlation bound follows by induction over independent Bernoulli coordinates. For bounded increasing \(F,G\), condition on the last coordinate \(Y\): the total covariance is \(\mathbb E\operatorname{Cov}(F,G\mid Y)+\operatorname{Cov}(\mathbb E[F\mid Y],\mathbb E[G\mid Y])\). The first term is nonnegative by induction; the second is \(p(1-p)\Delta F\Delta G\ge0\), where \(p=\Pr(Y=1)\) and \(\Delta F=\mathbb E[F\mid Y=1]-\mathbb E[F\mid Y=0]\), with \(\Delta G\) defined similarly. Negating one function gives the opposite-monotonicity form. The argument permits different parameters at different coordinates. The proof adapts the extremal-path resampling strategy of {[}BHK06{]} to connected interval errors. For a disconnected error closing \(\{-1,1\}\) with probability \(1/2\) while leaving 0 open, the two side-hole indicators have joint probability \(1/2\) and product of marginals \(1/4\). The conclusion is therefore false for arbitrary spatially correlated row errors. \subsubsection{A.2. Finite-history phase construction}\label{a.2.-finite-history-phase-construction} \textbf{Lemma A.2 (full insertion count).} In a translation-invariant hole field \(F\), insert \([0,r-1]\). Let \(L_z\) be the closed run immediately to the left of \(z\), and \(R_r\) the closed run beginning at \(r\). All indicator-weighted run expressions in this appendix are interpreted by restriction: \(\mathbf1_A Y\) equals \(Y\) on \(A\) and zero off \(A\), including when run lengths in \(Y\) are infinite off \(A\). In particular, each summand below is zero when \(z\in F\). The count is initially allowed to be infinite; under the hypotheses stated below, (A2) establishes its integrability. The number of newly closed, nonempty finite intervals is \[ \Delta_r(F)=\sum_{z=0}^{r-1}\mathbf{1}_{\{z\notin F\}}(L_z+1)(r-z+R_r). \tag{A1} \] Every newly closed interval has a unique first previously open point \(z\). Its left endpoint has \(L_z+1\) choices and its right endpoint has \(r-z+R_r\) choices. Assume \(u_F=\Pr(0\notin F)>0\) and the conditional inequality of Lemma A.1. If \(h_k(F)\) is the \(k\)-interval closure probability and \(\zeta=\lim_k h_k(F)\), boundary tail sums give \[ \mathbb E[\mathbf{1}_{\{z\notin F\}}(L_z+1)]=u_F+\beta_F-\zeta\le1. \] For \(d=r-z\), split \(\{z\notin F,[r,r+b-1]\subset F\}\) by the last open site in \([z,r-1]\), discard other restrictions, and sum over \(b\ge1\). Each of the \(d\) possible boundary sites contributes at most \(\beta_F-\zeta\), so \[ \mathbb E[\mathbf{1}_{\{z\notin F\}}R_r]\le d\beta_F,\qquad \mathbb E[\mathbf{1}_{\{z\notin F\}}(d+R_r)]\le d. \] Conditional negative association, followed by bounded truncation and monotone convergence, now yields \[ \mathbb E\Delta_r(F)\le V(r)/u_F. \tag{A2} \] \textbf{Lemma A.3 (finite-history bound).} If \(c\le1/32\), then \[ \sup_t\beta_t\le\frac{1-\sqrt{1-8c}}2<\frac12. \tag{A3} \] \textbf{Proof.} Use the last-row interpolation field \[ F_{t,\alpha}=E(H_t)\cup D^{(\alpha)},\qquad 0\le\alpha\le\varrho, \] where the first \(t\) rows have intensities \(\varrho\nu_r\) and the last row has intensities \(\alpha\nu_r\). It is exactly a finite-history process of Lemma A.1 with non-identical type parameters; its open density is at least \(u_{t+1}>0\). Work first on the infinite integer line with only lengths \(r\le K\) retained. This model remains translation invariant. A fixed finite observation and its derivative involve finitely many retained types in a backward cone of this spatially translation-invariant model. Define \(\phi_M(F,x)=\sum_{j=1}^M\mathbf{1}_{\{[x,x+j-1]\subset F\}}\). Poisson differentiation of the last row, followed by translation mass counting on this infinite-line law, gives \[ \frac{d}{d\alpha}\mathbb E\phi_M(F_{t,\alpha},0) =\sum_r\nu_r\,\mathbb E\Delta_{r,\le M}(F_{t,\alpha}) \le A_*/u_{t+1}. \tag{A4} \] Here \(\Delta_{r,\le M}\) counts new intervals of length at most \(M\). The same upper bound holds for all \(K\) since removal of long disasters only increases openness. Integrate over \(\alpha\) and let \(K\to\infty\) by bounded local convergence. Then let \(M\to\infty\) in the nonnegative insertion increments. Starting from zero, induction proves that \(\Phi_t:=\sum_{j\ge1}h_{j,t}<\infty\) for each finite \(t\): erosion does not increase it, and each row adds at most \(c/u_{t+1}\). Since erosion replaces \(h_{j,t}\) by \(h_{j+1,t}\), \[ \Phi_{t+1}-\Phi_t\le c/u_{t+1}-\beta_t. \tag{A5} \] Adding past rows makes the hole marginals larger, so \(\beta_t\) is nondecreasing. A single new row gives \(\beta_{t+1}\le\beta_t+q\) with \(q=1-e^{-\varrho S_1}\le c\). Independent time-block concatenation gives \(u_{a+b}\ge u_a u_b\): choose, by a fixed rule using only the last \(a\) rows, an open path to 0 and its initial site \(z\) when one exists. The first \(b\) rows reach that site with probability \(u_b\), independently. The shared boundary row is checked by the first block only. Thus \(\beta_{a+b}\le\beta_a+\beta_b\). Summing this inequality over complementary time indices, and summing (A5) while retaining \(\Phi_T\ge0\), gives \[ \frac{T-1}{2}\beta_T\le\sum_{t1/8\ge4c\), contradicting (A6). Hence the crossing never occurs. Taking \(T\to\infty\) in the sharper version of (A6) gives \(\beta(1-\beta)\le2c\) on the small branch, proving (A3). The constant \(1/32\) supplies a uniform positive-density regime. \(\square\) \subsubsection{A.3. Stationary flux}\label{a.3.-stationary-flux} The stationary field is a translation-equivariant measurable function of the independent columns of marked Poisson variables, hence is spatially ergodic. Equation (A3) implies \(u>0\), so it has no all-closed component and \(h_j\to0\). A stationary field cannot have a closed half-line with a finite endpoint at positive probability: the events specifying its unique endpoint are disjoint translates with equal probabilities. Thus closed runs are finite almost surely. The interpolation fields \(F_\alpha=E(H^*)\cup D^{(\alpha)}\) are coupled below \(F_\varrho\), which has law \(H^*\); they inherit this property and have open densities \(u_\alpha\ge u\). \textbf{Lemma A.4 (local stationary interpolation).} Let \(\Delta_{r,j}(F)\) count new closed intervals of length \(j\) after inserting \([0,r-1]\). Only the \(r+j-1\) windows intersecting the insertion can change, so \[ 0\le\Delta_{r,j}(F)\le r+j-1. \] For fixed \(j\), this bound is integrable under \(\nu\) by \(S_1<\infty\). Approximate \(H^*\) by finite histories and truncated lengths. Each local insertion difference converges and is bounded by the same integrable envelope. Dominated convergence passes the last-row interpolation to stationarity, giving \[ q_j:=h_j-h_{j+1}=\int_0^\varrho\sum_r\nu_r\mathbb E\Delta_{r,j}(F_\alpha)\,d\alpha. \tag{A7} \] The conditional inequality also passes to these fields from Lemma A.1, since their open densities are bounded below by \(u>0\). In a stationary field with \(h_j(F)\to0\), the boundary tail identity sharpens to \(\mathbb E[\mathbf{1}_{\{z\notin F\}}L_z]=\beta_F\). Substituting it in (A1) yields \[ \mathbb E\Delta_r(F)=V(r)+\sum_{z=0}^{r-1}\mathbb E[\mathbf{1}_{\{z\notin F\}}(L_z+1)R_r]\ge V(r). \] For the interpolation fields under consideration, \(h_j(F)\to0\). Combining the identity with (A2) and summing (A7) by Tonelli gives \[ \beta=\int_0^\varrho\sum_r\nu_r\mathbb E\Delta_r(F_\alpha)\,d\alpha, \qquad c\le\beta\le c/u. \tag{A8} \] The inequality \(\beta<1/2\) selects the small root of \(\beta(1-\beta)\le c\), giving \(\beta\le(1-\sqrt{1-4c})/2\). Also \(\beta\le2c\) and \(\beta-c\le\beta^2\le4c^2\). The phase bound (A3) determines the branch used in the stationary flux inequality. \subsubsection{A.4. Branching comparison and the spatial tail bound}\label{a.4.-branching-comparison-and-the-spatial-tail-bound} Put \(q_0=u\) and \(q_j=h_j-h_{j+1}\) for \(j\ge1\). This is a probability mass function on the nonnegative integers; let \(Q\) have this law, so \(\Pr(Q\ge k)=h_k\). For an interpolation field \(F\), define \(h^F_0=1\) and \(q^F_a=h^F_a-h^F_{a+1}\) for \(a\ge0\). This is also the probability that a site is open and the closed run immediately to its left has length at least \(a\). Assign each new interval of length \(j\) in (A1) to its first old open point \(z\). Let its left and right endpoint distances from \(z\) be \(a,b\) with \(a+b=j-1\), and put \(d=r-z\). If \(b0\), then: \begin{enumerate} \def\labelenumi{(\roman{enumi})} \item On \(O_z\), the two fields agree strictly left of \(z\). \item Given \(O_z\), increasing functions of constrained holes strictly left and strictly right of \(z\) are negatively correlated. \item If \(f\) is a bounded increasing function of unrestricted holes strictly left of \(z\), then \end{enumerate} \[ \mathbb E[f(H^0)\mid O_z]\le\mathbb E[f(H^0)\mid A_z]. \tag{B3} \] If \(Z\) is the rightmost constrained open site and \(\Pr(Z=z)>0\), then \[ \mathbb E[f(H^0)\mid Z=z]\le\mathbb E[f(H^0)\mid A_z]. \tag{B4} \] The statements pass to the infinite interval-type model at a fixed finite time whenever \(\sum_\ell\ell\theta_\ell<\infty\), and to increasing all-open prehistories when the limiting conditioning probabilities are positive. \textbf{Proof.} Work in a finite spatial slab containing every backward cone used in the observations. Initially take all type parameters strictly between zero and one. If an unrestricted open path ends at \(yK}(\ell+C)\theta_\ell\to0\). A feasible open path gives a positive lower bound for each conditioning probability as \(K\) increases. This passes bounded conditional expectations to the limit; deterministic parameters follow by parameter limits. Finally, extending an all-open prehistory gives monotone convergence of every finite terminal cylinder. A fixed checkpoint restricts only a finite cone for each terminal coordinate, so the same conditional limit works whenever the limiting event has positive probability. This proves the last assertion without a uniform history estimate. \(\square\) To justify finite endpoints in the stationary application, fix the number of rows after the checkpoint. Sufficiently far on its permitted side, every ancestor of a target is permitted, so constrained and unrestricted fields coincide there. The stationary unrestricted field has positive open density and is spatially ergodic (Appendix A). It has infinitely many open sites on that side. On the forbidden side there is a deterministic unreachable ray. The constrained boundary is therefore finite. Its event at any fixed location is a finite terminal cylinder, as required in Section 5.2. \subsection{Appendix C. Exact error-set boundary for the eroder comparison}\label{appendix-c.-exact-error-set-boundary-for-the-eroder-comparison} Gács and Törmä's \(\varepsilon\)-perturbation definition requires that, for every finite space-time set \(A\), the probability of a violation at every member of \(A\) is at most \(\varepsilon^{|A|}\) {[}GT22, Section 2{]}. It allows dependence; its restriction is on simultaneous errors. Choose an admissible positive background \(\theta_\ell=c\ell^{-6}\) with \(c>0\) small enough for the declared budgets. Start from all-open sites, so \(H_0=\varnothing\) and the actual rule-violation set at the first update is exactly \(D_1\). A length-\(k\) atom at origin zero closes \(k\) consecutive sites, giving \[ \Pr([0,k-1]\subset D_1)\ge1-e^{-ck^{-6}}. \tag{C1} \] For every fixed \(0<\varepsilon<1\), (C1) eventually exceeds \(\varepsilon^k\). Thus the common moment class does not satisfy that finite-error-set hypothesis. At the same time, (3.3) gives summable spatial pair covariances for this background. Summable pair dependence cannot substitute for an all-orders error-set condition. This proves the stated direct-application boundary; it does not assert failure of eroder stability under this noise. The independent space-time local-rule choices in {[}SST25, (1.7)--(1.8){]} give a second precise comparison. Our first-update local rules share interval atoms across sites, so their law lies outside that independent-rule set-up. The moment conditions in this paper control the recovery footprint of those correlated errors. \end{document}