\documentclass[10pt,a4paper]{article} \usepackage[margin=1in]{geometry} \usepackage[T1]{fontenc} \usepackage{lmodern,microtype,amsmath,amssymb,amsthm,mathtools} \usepackage{graphicx,booktabs,enumitem} \usepackage[numbers,sort&compress]{natbib} \usepackage[hidelinks]{hyperref} \usepackage{placeins} \hypersetup{pdftitle={Complete positivity and equilibrium realizations of reversible observations},pdfauthor={Samuil Petkov},pdfsubject={Finite equilibrium realization, reflected rank and copositive dissipation witnesses},pdfkeywords={equilibrium realization, detailed balance, hidden Markov models, reflection positivity, complete positivity, reflected rank, entropy production}} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{lemma}[theorem]{Lemma} \theoremstyle{remark}\newtheorem{remark}[theorem]{Remark} \newcommand{\CP}{\mathcal{CP}} \newcommand{\COP}{\mathcal{COP}} \newcommand{\E}{\mathbb E} \newcommand{\Pp}{\mathbb P} \newcommand{\one}{\mathbf 1} \newcommand{\KL}{D_{\mathrm{KL}}} \newcommand{\tv}{\mathrm{TV}} \newcommand{\sym}{\operatorname{sym}} \newcommand{\diag}{\operatorname{diag}} \newcommand{\spanof}{\operatorname{span}} \newcommand{\arcosh}{\operatorname{arcosh}} \newcommand{\artanh}{\operatorname{artanh}} \newcommand{\ip}[2]{\langle #1,#2\rangle} \newcommand{\pos}[1]{[#1]_+} \newcommand{\dc}{\Delta_{\mathrm{cp}}} \newcommand{\ff}{\mathcal F} \graphicspath{{figures/}} \title{Complete positivity and equilibrium realizations of reversible observations} \author{Samuil Petkov\\ \small Department of Physics, \'Ecole normale sup\'erieure, Universit\'e PSL, Paris, France\\ \small samuil.petkov@phys.ens.psl.eu} \date{5 September 2026} \begin{document} \maketitle \begin{abstract} Complete positivity of mirrored observation matrices is necessary for equilibrium realization, but does not ensure compatible equilibrium dynamics. We study stationary finite-state continuous-time Markov chains with instantaneous deterministic labels on a fixed physical clock. For reflection-positive laws with $d$ active labels and reflected rank $r\le d+1$, we construct a detailed-balance realization on the minimal number $r$ of states. Binary rank-four laws with two prediction dimensions per label compress to four states when a finite detailed-balance realization already exists. At $r=d+2$, a rational ten-state chain with six labels passes every mirrored matrix-CP test through one common nonnegative Gram factorization, yet admits no equilibrium realization. A rational seven-time statistic separates its law from all finite equilibrium observation laws, uniformly over hidden dimensions, rates and topology. It gives a positive entropy-production floor for every exact finite reciprocal realization, with conservative constants too small for practical statistical use. The law also has no ordinary symmetric Markov extension on any standard Borel state space at the original clock, despite factorability of its reflected half-path operator. For laws that fail a matrix-CP test, a separate deletion-deficit theorem gives a computable nonlinear dissipation bound. \end{abstract} \noindent\textbf{Mathematics Subject Classification (2020).} Primary 60J27; Secondary 82C05, 15B48. \noindent\textbf{Keywords.} equilibrium realization; detailed balance; hidden Markov models; reflection positivity; complete positivity; reflected rank; entropy production. \section{What reversible observations can establish about equilibrium} \label{sec:intro} An observed trajectory can be invariant under time reversal while its microscopic Markov mechanism violates detailed balance, as established for aggregated and hidden Markov models \citep{WagnerTimmer2000,Chen2011}. The non-uniqueness of aggregated Markov presentations of one observed law is classical \citep{FredkinRice1986,Kienker1989}. The realization question asks whether some finite detailed-balance mechanism produces exactly the observed law. An irreversible presentation alone cannot answer this question: the same law may also admit an equilibrium presentation. We study exact realization with instantaneous deterministic labels on a fixed continuous clock. At equilibrium, two observed histories read outward from a stationary midpoint are conditionally independent and identically distributed given the hidden midpoint. Every common finite classification of those histories therefore gives a completely positive matrix, a Gram sum of nonnegative vectors. This classical conditional-iid geometry \citep{SteinOzdaglarParrilo2013,BadescuODonnell2020} is stronger than entrywise nonnegative quadratic positivity \citep{Diananda1962,HallNewman1963,PfefferSamper2024}. It also connects the realization question to reflection-positive reconstruction. Reflection positivity supplies a self-adjoint observable semigroup \citep{Klein1978,JorgensenNeebOlafsson2016}, but does not generally supply a larger symmetric Markov process: Arveson's discrete-time nonextendable examples already establish that broader distinction \citep{Arveson1986}. A related but distinct finite-memory obstruction comes from \citet{FanizzaLumbrerasWinter2024}: some discrete-time processes with finite linear memory have no finite-dimensional quantum realization. There, complete positivity refers to quantum evolution maps; here it refers to entrywise nonnegative Gram factorizations of mirrored classical observation matrices. Neither representation contract settles whether the same continuous-time observed law has a finite ordinary detailed-balance realization with instantaneous labels and the original clock. The first result is an exact low-rank guarantee. Let $d$ be the number of active labels and $r$ the rank of the full reflected event kernel. Reflection positivity and $r\le d+1$ suffice for an irreducible detailed-balance realization on exactly $r$ states, preserving all labels and every cylinder on the original clock. The construction draws on established low-dimensional positive-realization and binary aggregated-model ideas \citep{Dharmadhikari1963,Heller1965,GrusslerDamm2012,WuJia2025,Siekmann2026} and enforces the label and generator constraints simultaneously. A complementary compression theorem treats binary rank four when each label has two prediction dimensions: any existing finite detailed-balance realization can be replaced by a four-state one, irrespective of its original hidden-state count. The two replacements must be selected jointly to preserve nonnegative cross rates. The general low-rank guarantee nevertheless ends at $r=d+1$. A rational ten-state model with six labels and $r=d+2$ passes every mirrored matrix-CP test using one common nonnegative factorization, but has no finite detailed-balance realization of its complete observed law. A seven-time rational statistic separates it from all finite equilibrium observations and certifies a positive dissipation floor across exact compatible models. The obstruction concerns symmetric nonnegative dynamics, rather than incompatible choices of Gram factors for separate windows. For thermodynamic inference, a violation of one CP test has a quantitative consequence. Observed path-relative entropy vanishes for a reversible law \citep{Roldan2010,Seifert2012}, while counts, waiting times and constraints on microscopic statistics can retain information \citep{Skinner2021,SkinnerWaiting2021,Nitzan2023,AguileraItoKolchinsky2026}. Equilibrium restrictions on binary dwell laws already provide nonequilibrium tests \citep{Tu2008}; Skinner and Dunkel also give reversible observed trajectories with positive certified hidden dissipation \citep{Skinner2021}. Thus reversible observations concealing dissipation are a precedent, not the novelty claim here. Hidden entropy production under coarse graining and for a marginal observer has been treated in stochastic thermodynamics \citep{Esposito2012,PolettiniEsposito2017}, and fits of hidden Markov models to observed data give lower bounds on it \citep{Ehrich2021}. Transition-indexed waiting-time methods \citep{vanderMeer2022,FritzSeifert2026} and recently resolved excursion or cycle methods \citep{Gu2026,Biddle2026} use complementary observation contracts. Here the completely positive deletion deficit of an arbitrary matched classifier gives a nonlinear entropy-production bound independent of the finite hidden dimension and topology. Its convex weight and copositive dual have classical geometric antecedents \citep{DucuaraSkrzypczyk2020}; the contribution is their link to microscopic Markov path divergence. Finite-observable certificates for dissipation plus a positive price per hidden state require global model calibration \citep{Aznagulov2026Architecture}. Our explicit separation concerns the unpenalized union of finite equilibrium architectures. Section~\ref{sec:model} fixes the observation contract. Section~\ref{sec:realization} states the low-rank realization and binary compression theorems, followed by the all-window-CP counterexample. Section~\ref{sec:main} gives the finite-window dissipation theorem, followed by one reversible benchmark and fixed-rule sampling guarantees. All full proofs, the rational counterexample and the additional robustness and kinetic comparisons are included after the references. Exact finite equilibrium nonexistence and a positive dissipation infimum remain distinct throughout. \emph{Scope.} Classical ingredients are credited to the cited sources, and all results are stated for the observation contract of Section~\ref{sec:model} only. \section{Observation model and physical conventions} \label{sec:model} Let $X$ be a finite, irreducible, stationary continuous-time Markov chain with row generator $Q$, stationary law $\pi$ and reciprocal support: $q_{xy}>0$ if and only if $q_{yx}>0$. Missing reciprocal pairs are allowed. Every microscopic state is even under physical time reversal. Write \begin{equation} \sigma(Q)=\sum_{x0$, his conditional word rank $n_\Delta(a)$ equals $\dim\mathcal H_a$ here, as proved in Appendix~\ref{app:rank-dictionary}. Thus the ordinary $r$-state discrete-time realization count on this slice is already supplied by that corollary. Invariant polyhedral cones are classical realization data \citep{Heller1965,Fliess1975}. Symmetric scalar positive realization is treated by \citet{GrusslerDamm2012}. For binary three-state aggregated models, an explicit rotation to a detailed-balance equivalent appears in the author preprint of \citet{WuJia2025}, Appendix~A; related binary equivalence classes are analyzed by \citet{Siekmann2026}. Theorem~\ref{thm:rank} concerns simultaneous preservation of every label projection and every continuous-time cylinder. It requires no renewal assumption. For example, a four-state equilibrium chain with labels $(a,a,b,c)$ can have rank four and entrance-dependent visits to $a$; an exact example is given at the end of Appendix~\ref{app:rank}. The binary case with two two-dimensional label spaces admits a further compression when an equilibrium presentation is already available. The space $V$ below is the span of conditional future-cylinder functions; at equilibrium it is isometric to the reflected space $\mathcal H$. \begin{theorem}[Minimal binary equilibrium compression] \label{thm:binary-compression} Let $Q$ be a finite irreducible detailed-balance CTMC with two nonempty instantaneous deterministic labels $a,b$, and let $V$ be the smallest real space of state functions containing $\one$ and invariant under $Q$ and the two label masks $\one_a,\one_b$. If \begin{equation} \dim V=4,\qquad \dim(\one_aV)=\dim(\one_bV)=2, \label{eq:binary-rank-profile} \end{equation} then its stationary observed law has an exact irreducible four-state detailed-balance realization, with two states per label and the same physical clock. Four states are minimal among all finite CTMC realizations of that law. The original number of states in either label is unrestricted. \end{theorem} Each label's prediction coordinate can be normalized to have conditional mean zero and variance one. Detailed balance then expresses the cross-label action by a bilinear rate kernel. Conditional destination means make this kernel nonnegative on two crossing strips. Since its value at the origin is positive, one outer corner is positive. Retaining that corner and replacing the opposite coordinates by their negative reciprocals supplies two two-point measures with matching moments and nonnegative cross rates. The retained extrema also determine the signs of both internal rates. Appendix~\ref{app:binary-compression} verifies the generator, full cylinder law, irreducibility and minimality. Dharmadhikari's conditional-rank corollary cited above supplies the ordinary four-state realization count for each sampled binary law; Theorem~\ref{thm:binary-compression} preserves detailed balance and the entire continuous clock in one realization. Its assumption of an existing equilibrium presentation is essential to the stated compression result. Every equilibrium realization passes all mirrored matrix-CP tests. Theorem~\ref{thm:rank} shows that reflection positivity alone suffices at $r\le d+1$. The next result places a limitation exactly one rank increment beyond that general sufficient regime. \begin{theorem}[All-window complete positivity without a finite equilibrium lift] \label{thm:allcp} There exists a ten-state irreducible stationary CTMC with rational rates, reciprocal support and six instantaneous deterministic labels whose full observed law is reversible and has reflected rank eight. For every positive window and every finite measurable matched classifier, its mirrored matrix is completely positive, with a common factorization through the ten microscopic coordinates. Nevertheless, no finite stationary detailed-balance CTMC with instantaneous deterministic labels has that same full observed law on the same clock, at any hidden-state count. \end{theorem} The construction uses five core states sharing one label and five distinct probe labels. Core states carry unit-circle prediction points $p_i$. Nonnegative affine probe rates cut out their polygon. In any hypothetical equilibrium realization, the same observed law forces every competing core state into that polygon. Its second moment saturates the unit-radius bound, so all these states lie at the original vertices, with the original aggregate masses. Detailed balance would then impose symmetric nonnegative conductances between the vertices. A rational separating identity rules them out. Appendix~\ref{app:allcp} supplies the full generator, the common all-window factorization and the argument over every competing state count. The example has one three-dimensional label space and five one-dimensional spaces: $r=d+2$. Discrete-time reflection-positive processes without symmetric Markov extensions were already constructed by \citet{Arveson1986}. His extension contract allows arbitrary standard Borel state spaces, and his obstruction uses nonfactorability of the reflected half-path operator. Theorem~\ref{thm:allcp} is a finite continuous-time construction satisfying the stronger all-window matrix-CP condition. The connection can be made exact: for the observed half-path $L^2$ space, all these partition matrices are CP if and only if its reflected operator $\mathsf K$ is factorable as $B^*B$ with $B$ order-preserving. Appendix~\ref{app:extensions} proves this comparison using Arveson's closed-cone theorem. For the construction above, $B$ is simply conditional future expectation given the original ten-state midpoint. The continuous-time symmetric-extension setting was subsequently discussed by Mitter \citep[pp.~104--105]{Mitter1991}; the present observed finite-CTMC law is continuous in probability. We retain the explicit standard Borel, ordinary-symmetry and deterministic-observation contract below. The historical formulation alone is not a present-day priority determination. Ordinary symmetry leaves microscopic state values unchanged under reversal; no additional odd-variable involution is allowed. \begin{corollary}[No ordinary symmetric same-clock extension] \label{cor:general-extension} The observed law in Theorem~\ref{thm:allcp} has no realization as an instantaneous deterministic Borel observation of a stationary two-sided ordinary symmetric Markov process on a standard Borel state space, at the same continuous clock. \end{corollary} The proof extends the polygon rigidity to measurable prediction points. Saturation leaves only ten observable prediction classes. A finite-valued conductance lemma then reproduces the finite observable restriction by a finite detailed-balance generator, contradicting the same rational certificate. The observed law supplies an analytic semigroup on its finite-dimensional observable image; no global microscopic continuity, jump kernel, finite rate, or generator-domain condition on the class indicators is assumed. This conclusion is a continuous-clock extension statement; it does not settle the extension problem for an isolated sampled discrete-time process. Exact equilibrium nonexistence alone does not establish a positive dissipation infimum or resistance to approximation. For this law, a finite-observation certificate gives both conclusions explicitly. Let $\mathcal E_{\mathrm{fin}}$ contain all stationary finite detailed-balance CTMC observation laws on these six labels, with unrestricted hidden dimensions, rates and topology; reducible models are allowed. Enumerate finite strictly increasing rational time tuples containing zero as $(\mathcal T_m)_{m\ge1}$ and set \begin{equation} D(P,P')=\sum_{m\ge1}2^{-m}\tv(P_{\mathcal T_m},P'_{\mathcal T_m}). \label{eq:cylinder-metric-main} \end{equation} \begin{theorem}[An explicit certificate beyond matrix-CP tests] \label{thm:closure} For the law $P$ in Theorem~\ref{thm:allcp}, put $h=10^{-3}$, $t=10^{-7}$ and \begin{equation} \mathcal T=(0,t,h,h+t,h+2t,2h+t,2h+2t),\qquad T=2(h+t). \label{eq:certificate-times-main} \end{equation} The rational word statistic $W\colon\{a,b_1,\ldots,b_5\}^7\to[0,1]$ specified in \eqref{eq:certificate-score}--\eqref{eq:certificate-normalization}, together with the constants $c=1/2+1/(192B)$ and $\eta=71/(4800B)>0$ defined there, satisfies \begin{equation} \E_PW\ge c+\eta,\qquad \sup_{P'\in\mathcal E_{\mathrm{fin}}}\E_{P'}W\le c. \label{eq:certificate-separation-main} \end{equation} Consequently, if $m_0$ indexes $\mathcal T$ in \eqref{eq:cylinder-metric-main}, \begin{equation} \inf_{P'\in\mathcal E_{\mathrm{fin}}}D(P,P')\ge2^{-m_0}\eta>0, \qquad I_\infty(P):=\inf_{Q\text{ exactly compatible}}\sigma(Q) \ge\frac{8\eta^2}{T}>0. \label{eq:robust-positive-floor} \end{equation} The second infimum ranges over all finite irreducible stationary reciprocal CTMCs with instantaneous deterministic labels producing exactly $P$ on the same clock. \end{theorem} The score tests whether predictions near the polygon vertices can evolve consistently with a prescribed rational finite-lag action. Every mismatch of an unknown competitor with this action enters an observed residual penalty; no model fit or rate bound is assumed. A first-moment polygon estimate controls rounding errors using a weakly monotone dual field, including its zero-coercivity pairs. Appendix~\ref{app:closure} gives the construction and the reversibilization argument for the entropy-production bound. The margin is approximately $1.2398\times10^{-26}$, so the certificate establishes effective separation without practical statistical power. Thus all matrix-CP deficits can vanish while a finite observed statistic certifies a positive hidden-dissipation floor. The theorem below instead applies when a tested matrix itself fails CP. \section{A population bound for arbitrary finite classifiers} \label{sec:main} Fix a half-window $\tau>0$ and one measurable finite classifier $\Phi$. Write $Y^{\leftarrow}_0=Y_0$ and $Y^{\leftarrow}_s=h(X_{(-s)-})$ for $s>0$, using the usual right-continuous reversal. Deterministic endpoints are almost surely not jump times. Define \begin{equation} A=\Phi((Y^{\leftarrow}_s)_{0\le s\le\tau}),\qquad B=\Phi((Y_s)_{0\le s\le\tau}),\qquad M_{ij}=\Pp(A=i,B=j),\quad S=\frac{M+M^{\mathsf T}}2. \label{eq:motif} \end{equation} The past is read \emph{outward} from the midpoint, using the same rule as the future. The full record has duration $2\tau$. A common stochastic classifier is also allowed when independent auxiliary randomness is used on the two halves. Arbitrary correlated detector memory is not covered by this observation contract. This distinction matters for entropy inference from last-visited records \citep{Martinez2019,HartichGodec2024,BiskerReply2024}. For $m$ categories, let $\CP_m=\{\sum_r u_ru_r^{\mathsf T}:u_r\ge0\}$ and $\COP_m=\{W=W^{\mathsf T}:u^{\mathsf T}Wu\ge0\ \forall u\ge0\}$. Pairings sum over \emph{ordered} indices. Set \begin{align} \dc(S)&=1-\max\{\one^{\mathsf T}G\one:G\in\CP_m,\ 0\le G\le S\}, \label{eq:deletion}\\ \ff(v)&=\sqrt v\,\artanh\sqrt v =\sum_{k=1}^{\infty}\frac{v^k}{2k-1},\quad 0\le v<1. \label{eq:F} \end{align} All inequalities in \eqref{eq:deletion} are entrywise. The retained mass $1-\dc(S)$ is the largest completely positive fraction of $S$; equivalently, $\dc$ is the convex resource weight with normalized completely positive matrices as the free set \citep{DucuaraSkrzypczyk2020}. The dual representation below is this classical weight geometry specialized to categorical pair distributions. General copositive optimization need not be computationally tractable. \begin{theorem}[Finite-window dissipation witness] \label{thm:main} Under the finite reciprocal CTMC observation contract of Section~\ref{sec:model}, without requiring observed reversibility, let $\sigma(Q)$ be as in \eqref{eq:sigma}, $M,S$ as in \eqref{eq:motif}, and $\dc,\ff$ as in \eqref{eq:deletion}--\eqref{eq:F}. Then \begin{equation} \boxed{\quad \sigma(Q)\ \ge\ \frac2\tau\ff(\dc(S)).\quad} \label{eq:main} \end{equation} In particular, for any symmetric copositive $W$ with $W_{ij}\ge-1$, put $v_W=\pos{-\ip W M}$. Then $0\le v_W\le\dc(S)<1$ and \begin{equation} \sigma(Q)\ge\frac2\tau\ff(v_W)\ge\frac{2v_W}{\tau}. \label{eq:dualbound} \end{equation} At equilibrium $M$ is completely positive. Conversely, $\dc(S)=0$ holds exactly when $S$ is completely positive; this is a statement about one tested matrix, not a characterization of the entire observed process. \end{theorem} \begin{proof}[Proof of the thermodynamic bound] Condition on $X_0=x$ and write $a_{xi}=\Pp(A=i\mid X_0=x)$ and $b_{xi}=\Pp(B=i\mid X_0=x)$. Conditional independence gives $M=\sum_x\pi_xa_xb_x^{\mathsf T}$. Define \begin{equation} z_{xi}=\sqrt{a_{xi}b_{xi}},\qquad G=\sum_x\pi_x z_xz_x^{\mathsf T},\qquad H_x=\sum_i\sqrt{a_{xi}b_{xi}}. \label{eq:gram} \end{equation} The arithmetic--geometric mean inequality gives $0\le G\le S$, and $G$ is completely positive. Its diagonal equals that of $S$. Thus \begin{equation} \dc(S)\le \delta:=1-\one^{\mathsf T}G\one =\sum_x\pi_x(1-H_x^2). \label{eq:delta} \end{equation} The elementary relative-entropy/affinity inequality, proved in Appendix~\ref{app:affinity}, is \begin{equation} J(p,q):=\KL(p\Vert q)+\KL(q\Vert p) \ge4\ff(1-H(p,q)^2). \label{eq:Jeffreys} \end{equation} Convexity of $\ff$ then gives $\sum_x\pi_xJ(a_x,b_x)\ge4\ff(\delta)$. The stationary reverse generator is $q^\dagger_{xy}=\pi_yq_{yx}/\pi_x$. Past half-paths conditioned on $x$ have its forward law. The two averaged conditional path divergences, before classification, are each $\tau\sigma(Q)$. Applying data processing in both directions yields \begin{equation} 4\ff(\dc(S))\le4\ff(\delta) \le\sum_x\pi_xJ(a_x,b_x)\le2\tau\sigma(Q). \label{eq:chain} \end{equation} The path identities, support issues and coefficient $2/\tau$ are justified in Appendix~\ref{app:path}. For a copositive $W\ge-1$, $-\ip W S\le\one^{\mathsf T}(S-G')\one$ for every feasible $G'$ in \eqref{eq:deletion}. This proves \eqref{eq:dualbound}. At equilibrium $a_x=b_x$, yielding the claimed factorization of $M$. The remaining cone assertions follow from Proposition~\ref{prop:duality}. \end{proof} \begin{proposition}[Optimized witnesses and coarsening] \label{prop:duality} For any symmetric nonnegative $S$ of total mass one, \begin{equation} \dc(S)=\sup\{-\ip W S:W\in\COP_m,\ W_{ij}\ge-1\}. \label{eq:duality} \end{equation} The primal maximum exists and has $\diag G=\diag S$. If $K$ is any nonnegative row-stochastic matrix, then \begin{equation} \dc(K^{\mathsf T}SK)\le\dc(S). \label{eq:contraction} \end{equation} Consequently refining the same classifier on both halves cannot decrease the optimized deficit at fixed $\tau$; independent matched relabeling cannot increase it. \end{proposition} The proof in Appendix~\ref{app:cone} uses only finite-dimensional separation and nonnegative Gram factors. A zero deficit is not a certificate of an equilibrium Markov realization. In particular, compatible factorizations at separate window lengths need not form one Markov semigroup. For overlapping relabeling channels the contraction can be quantified: \begin{equation} \dc(K^{\mathsf T}SK)\le c(K)\dc(S),\qquad c(K)=1-\min_{i,j}\left(\sum_a\sqrt{K_{ia}K_{ja}}\right)^2. \label{eq:channel-main} \end{equation} The coefficient is sharp over all symmetric nonnegative probability matrices, as proved in Appendix~\ref{app:cone}. Sharpness over physically compatible CTMC motif matrices is a separate question. Leaving a censoring class isolated can give $c(K)=1$ even when active classes overlap. \section{Five motif classes and an explicit reversible benchmark} \label{sec:example} Arrange five active motif classes on the cycle $C_5$, and reserve class $0$ for censoring or nonselection. With $A_5$ its adjacency matrix, let $W=J_5-2A_5$ on the active classes and put zero in its censoring row and column. This is the classical Horn copositive witness. A short proof is given in Appendix~\ref{app:horn}. Define \begin{equation} Z=2\mathbf1_{\{A\sim B\}}-\mathbf1_{\{A\ne0,\,B\ne0\}}, \qquad v=\pos{\E Z}. \label{eq:score} \end{equation} Here adjacency excludes equal classes. The observable score takes values in $\{-1,0,1\}$, and \begin{equation} \boxed{\quad\sigma\ge\frac2\tau\ff(v)\ge\frac{2v}{\tau}.\quad} \label{eq:cyclebound} \end{equation} At equilibrium, conditional on both classes being active, their adjacency probability cannot exceed $1/2$. This implication is used only when the conditioning event has positive probability; censored records are retained in the unconditional score. \subsection{Calibration against an attained full-law optimum} The inherited construction \citep{PetkovHidden} has five visible states, one common label $H$ for twenty hidden states, and \begin{equation} R=I+\frac35 A_5+\frac1{100}(J_5-I-A_5),\qquad f(t)=\lambda e^{-\lambda t}R,\qquad \lambda=\frac{111}{5}. \label{eq:fixture} \end{equation} Visible stationary masses are $1/10$, total hidden mass is $1/2$, and the symmetric direct visible flows are $F_{ij}=1/100$ for $i\ne j$. Here $R_{ij}$ is the stationary occurrence rate of an excursion from visible state $i$ through $H$ to visible state $j$, including returns. The full observed stationary process is reversible. Its exact minimum entropy production among finite reciprocal realizations is \begin{equation} I_{\infty}^{\mathrm{full}}= \frac{3\sqrt{111}}5\arcosh\frac{20}{17} \simeq3.70232333. \label{eq:exact} \end{equation} Appendix~\ref{app:benchmark} supplies the microscopic channels, stationarity, the inherited sharp lower bound and attainment. This value is used for calibration; it is \emph{not} claimed to be determined by a single motif matrix. For a half-record starting in $H$, let $\Phi$ return the first visible state reached by time $\tau$. Return $0$ if no visible state is reached, or if the midpoint starts in a visible state. For $q=e^{-\lambda\tau}$ and $t_\tau=(1-q)^2/\lambda$, direct stationary calculation gives \begin{equation} M=\begin{pmatrix} (1+q^2)/2 & q(1-q)\one^{\mathsf T}/10\\ q(1-q)\one/10 & t_\tau R \end{pmatrix},\qquad v=\frac9{10}t_\tau. \label{eq:M} \end{equation} The whole six-category matrix is positive definite, not just its active principal block. Its optimized deficit equals this same score: $\dc(M)=v$, by an explicit primal Gram certificate in Appendix~\ref{app:benchmark}. The linear bound is largest at $x=\lambda\tau$ solving $e^x=1+2x$, with $x\simeq1.25643121$. It gives $0.73307588$ in the chosen inverse-time unit. At the same window the nonlinear bound is $0.73820904$, approximately $19.9\%$ of \eqref{eq:exact}. The nonlinear optimum is not asserted to occur at the linear optimizer. These decimals are numerical evaluations of proved formulas. \subsection{A finite noise tolerance} Independently on each half, keep an active class with probability $1-\eta$ and replace it by a uniform class in $\{1,\ldots,5\}$ with probability $\eta$. Class $0$ is unchanged. This is the same stochastic map in both directions, so it preserves the validity of the thermodynamic test. The new score is exactly \begin{equation} \E Z_\eta=t_\tau \left[\frac{78}{25}(1-\eta)^2-\frac{111}{50}\right]. \label{eq:noise} \end{equation} It is positive precisely for \begin{equation} \eta<\eta_*:=1-\sqrt{\frac{37}{52}}\simeq0.15647261. \label{eq:threshold} \end{equation} In fact, the whole noisy matrix has $\dc(M_\eta)=\pos{\E Z_\eta}$: it enters the completely positive cone exactly at $\eta_*$. Appendix~\ref{app:benchmark} gives a nonnegative Gram certificate on both sides of this threshold. Complete positivity at one window does not imply equilibrium realizability of the full process. This is a population robustness threshold for \emph{uniform motif relabeling}, not a universal detector-error tolerance. It does not cover asymmetric filters or arbitrary errors in transition times. Both reversibility and positive semidefiniteness of the matrix are preserved by this matched corruption. \begin{figure}[t] \centering\includegraphics[width=.84\linewidth]{noise_tolerance.pdf} \caption{Normalized mean witness under independent uniform relabeling of the active motif categories, at any fixed positive half-window. The exact zero is $1-\sqrt{37/52}$. Relabeling never creates a violation for an equilibrium process; here it gradually removes the available signal.} \label{fig:noise} \end{figure} The benchmark can also be extended to a reciprocal 50-state copy-switch model in which every observed label is many-to-one. Its full observed law remains reversible and its tested matrix remains positive definite with a strictly positive CP witness for every finite positive switching vector. Appendix~\ref{app:lift} gives the construction and exact score. This is a robustness extension of the non-CP benchmark, separate from the all-window-CP counterexample of Theorem~\ref{thm:allcp}. A four-stage driven cycle supplies a second check of the finite-window certificate. With two compound labels, it has a reversible observed law and a quadratic weak-driving signal, whereas the specified persistence mutual-information certificate of \citet{AnonymousHiddenDissipation2026} is quartic on the same data. This binary example tests estimator strength, rather than CP versus PSD. The exact formulas, heterogeneous-rate radius and comparison are in Appendix~\ref{app:catalytic-baseline}. \section{Finite-data statements and their assumptions} \label{sec:statistics} Population inequalities and empirical confidence bounds are different assertions. Fix $\Phi$, $\tau$ and the cycle assignment in advance, or choose them using training records independent of the test records. For $n$ independent stationary snippets, let $\overline Z_n$ be the average score. Hoeffding's inequality \citep{Hoeffding1963} gives, with probability at least $1-\alpha$, \begin{equation} \sigma\ge\frac2\tau\ff\left( \pos{\overline Z_n-\sqrt{2\log(1/\alpha)/n}}\right). \label{eq:iid} \end{equation} The three-valued score also permits a sharper construction using classical exact binomial confidence limits \citep{ClopperPearson1934}. Let $N$ be the number of nonzero scores and $K$ the number of positive scores. Write $L(k,n;\gamma)$ for the one-sided exact binomial lower endpoint, defined in Appendix~\ref{app:stats}. With probability at least $1-\alpha$, \begin{equation} \sigma\ge\frac2\tau\ff\!\left( L(N,n;\alpha/2)\, \pos{2L(K,N;\alpha/2)-1}\right). \label{eq:binomialconfidence} \end{equation} An exact equilibrium test follows by conditioning on $N$: under equilibrium $K$ has law $\operatorname{Bin}(N,p)$ with $p\le1/2$. Its upper tail at $p=1/2$ is a valid conservative $p$-value. This is a test of the specified equilibrium model class, not a diagnosis of which microscopic assumption fails. For a single stationary trajectory, the same event-transfer proof applies to nonoverlapping snippets separated by gaps only when an external hidden-state mixing bound is justified. With the bound $b(g)$ in \eqref{eq:mix}, the independent-snippet failure probability increases by at most $(n-1)b(g)$. Appendix~\ref{app:stats} gives this assumption, training and candidate-selection corrections, exact binomial power and contamination bounds. Empirical autocorrelation is not substituted for a hidden-state mixing guarantee. \section{Interpretation and remaining scope} Mirrored observed histories distinguish matrix positivity, equilibrium realization and dissipation. A non-CP tested matrix certifies a uniform positive entropy-production floor for compatible finite reciprocal mechanisms. At reflected rank $r\le d+1$, reflection positivity gives a minimal exact finite equilibrium realization. Beyond that regime, even a common CP hierarchy need not admit dynamically compatible equilibrium rates. The rational example makes the rank-gap guarantee sharp in general, without classifying every alphabet/rank pair or label-block geometry. The realization and thermodynamic conclusions have different quantifiers. Theorem~\ref{thm:allcp} excludes every exact finite ordinary detailed-balance CTMC; Corollary~\ref{cor:general-extension} also excludes arbitrary-state stationary ordinary symmetric Markov extensions. The finite-word separation in Theorem~\ref{thm:closure}, rather than exact nonexistence alone, supplies an explicit positive entropy-production floor over all exact compatible finite reciprocal models. Its very small margin does not provide a practical sample-size guarantee. Theorem~\ref{thm:main} gives a different positive floor when a finite observed matrix fails CP; neither bound is asserted to attain the full-law dissipation optimum. All results retain instantaneous labels, the original clock and even microscopic time-reversal parity. Correlated detector memory and generalized odd-variable reversal require different contracts. \paragraph{Reproducibility.} The appendices give complete proofs and exact rational specifications of the generators, observable spaces, facet and conductance certificates, finite-window formulas and confidence calculations. Numerical evaluations illustrate analytical results and are not proofs of the all-time or arbitrary-state-count claims. The ancillary file \texttt{anc/check\_certificate.py} (Python standard library only) recomputes, in exact rational arithmetic and, for the target bounds, outward-rounded rational interval arithmetic, the data of Theorem~\ref{thm:allcp} and every numerical input of the certificate in Theorem~\ref{thm:closure}, including the target bounds \eqref{eq:certificate-target-moments}. \section*{AI assistance disclosure} OpenAI GPT-5.6 Sol, accessed through ChatGPT and Codex, was used to develop and audit the proofs, exact certificates and numerical experiments, and for drafting. Anthropic Claude Opus 5.5, used through Claude Code, was used for editing, an exact-arithmetic verification of the certificate in Theorem~\ref{thm:closure}, the improved constants, and the ancillary checker. The author reviewed and takes responsibility for every statement and proof. AI systems are not authors of this work, and the work has not been peer reviewed. \FloatBarrier \begin{thebibliography}{99} \bibitem{Roldan2010} E. Rold\'an and J. M. R. Parrondo, Estimating dissipation from single stationary trajectories, \emph{Phys. Rev. Lett.} \textbf{105}, 150607 (2010), doi:10.1103/PhysRevLett.105.150607. \bibitem{Seifert2012} U. Seifert, Stochastic thermodynamics, fluctuation theorems and molecular machines, \emph{Rep. Prog. Phys.} \textbf{75}, 126001 (2012), doi:10.1088/0034-4885/75/12/126001. \bibitem{Skinner2021} D. J. Skinner and J. Dunkel, Improved bounds on entropy production in living systems, \emph{Proc. Natl. Acad. Sci. USA} \textbf{118}, e2024300118 (2021), doi:10.1073/pnas.2024300118. \bibitem{SkinnerWaiting2021} D. J. Skinner and J. Dunkel, Estimating entropy production from waiting time distributions, \emph{Phys. Rev. Lett.} \textbf{127}, 198101 (2021), doi:10.1103/PhysRevLett.127.198101. \bibitem{Nitzan2023} E. Nitzan, A. Ghosal and G. Bisker, Universal bounds on entropy production inferred from observed statistics, \emph{Phys. Rev. Research} \textbf{5}, 043251 (2023), doi:10.1103/PhysRevResearch.5.043251. \bibitem{vanderMeer2022} J. van der Meer, B. Ertel and U. Seifert, Thermodynamic inference in partially accessible Markov networks: A unifying perspective from transition-based waiting time distributions, \emph{Phys. Rev. X} \textbf{12}, 031025 (2022), doi:10.1103/PhysRevX.12.031025. \bibitem{Diananda1962} P. H. Diananda, On non-negative forms in real variables some or all of which are non-negative, \emph{Math. Proc. Cambridge Philos. Soc.} \textbf{58}, 17--25 (1962), doi:10.1017/S0305004100036185. \bibitem{HallNewman1963} M. Hall, Jr. and M. Newman, Copositive and completely positive quadratic forms, \emph{Math. Proc. Cambridge Philos. Soc.} \textbf{59}, 329--339 (1963), doi:10.1017/S0305004100036951. \bibitem{PfefferSamper2024} M. Pfeffer and J. A. Samper, The cone of $5\times5$ completely positive matrices, \emph{Discrete Comput. Geom.} \textbf{71}, 442--466 (2024), doi:10.1007/s00454-023-00620-y. \bibitem{PetkovHidden} S. Petkov, \emph{Entropy Production under Hidden-State Reduction: Exact Safety and Statistical Limits}, with Technical Supplement, unpublished manuscript, 4 October 2026. The inherited result is the supplement's exact five-cycle dissipation theorem; Appendix~\ref{app:benchmark} reproduces a self-contained proof of the comparison value used here. \bibitem{Hoeffding1963} W. Hoeffding, Probability inequalities for sums of bounded random variables, \emph{J. Amer. Statist. Assoc.} \textbf{58}, 13--30 (1963), doi:10.1080/01621459.1963.10500830. \bibitem{ClopperPearson1934} C. J. Clopper and E. S. Pearson, The use of confidence or fiducial limits illustrated in the case of the binomial, \emph{Biometrika} \textbf{26}, 404--413 (1934), doi:10.1093/biomet/26.4.404. \bibitem{MotzkinStraus1965} T. S. Motzkin and E. G. Straus, Maxima for graphs and a new proof of a theorem of Tur\'an, \emph{Canad. J. Math.} \textbf{17}, 533--540 (1965), doi:10.4153/CJM-1965-053-6. \bibitem[Aguilera et~al.(2026)]{AguileraItoKolchinsky2026} M.~Aguilera, S.~Ito, and A.~Kolchinsky, \emph{Inferring Entropy Production in Many-Body Systems Using Nonequilibrium Maximum Entropy}, Physical Review Letters \textbf{136}, 077101 (2026), \href{https://doi.org/10.1103/xgkj-dxzh}{doi:10.1103/xgkj-dxzh}. \bibitem[Anonymous(2026)]{AnonymousHiddenDissipation2026} Anonymous, \emph{An architecture-independent dissipation floor for alternating-renewal observations}, Zenodo preprint (unrefereed), version 1.1.0-candidate, 3 October 2026, \href{https://doi.org/10.5281/zenodo.23118474}{doi:10.5281/zenodo.23118474}. The fixed-window Corollary~7 does not require renewal. \bibitem[Fritz and Seifert(2026)]{FritzSeifert2026} J.~H.~Fritz and U.~Seifert, \emph{Waiting-time-based entropy estimators in continuous space without Markovian events}, Physical Review E \textbf{114}, 034109 (2026), \href{https://doi.org/10.1103/zj2b-w93z}{doi:10.1103/zj2b-w93z}. \bibitem[Stein et~al.(2013)]{SteinOzdaglarParrilo2013} N.~D.~Stein, A.~Ozdaglar, and P.~A.~Parrilo, \emph{Exchangeable Equilibria, Part I: Symmetric Bimatrix Games}, \href{https://arxiv.org/abs/1307.3586}{arXiv:1307.3586} (2013). \bibitem[B\u{a}descu and O'Donnell(2020)]{BadescuODonnell2020} C.~B\u{a}descu and R.~O'Donnell, \emph{Lower Bounds for Testing Complete Positivity and Quantum Separability}, in \emph{LATIN 2020: Theoretical Informatics}, Lecture Notes in Computer Science \textbf{12118}, 375--386 (Springer, 2020), \href{https://doi.org/10.1007/978-3-030-61792-9_30}{doi:10.1007/978-3-030-61792-9\_30}. \bibitem[Ducuara and Skrzypczyk(2020)]{DucuaraSkrzypczyk2020} A.~F.~Ducuara and P.~Skrzypczyk, \emph{Operational Interpretation of Weight-Based Resource Quantifiers in Convex Quantum Resource Theories}, Physical Review Letters \textbf{125}, 110401 (2020), \href{https://doi.org/10.1103/PhysRevLett.125.110401}{doi:10.1103/PhysRevLett.125.110401}. \bibitem[Raux et~al.(2026)]{Raux2026} P.~Raux, A.~M.~Maier, and U.~Seifert, \emph{Lower bounds on entropy production from dynamical correlation functions}, \href{https://arxiv.org/abs/2608.03619}{arXiv:2608.03619} (2026). \bibitem{ChazottesGiardinaRedig2005} J.-R.~Chazottes, C.~Giardina and F.~Redig, Relative entropy and waiting times for continuous-time Markov processes, \href{https://arxiv.org/abs/math/0512386v1}{arXiv:math/0512386v1} (2005). \bibitem{Mitter1991} S.~K.~Mitter, Markov Random Fields, Stochastic Quantization and Image Analysis, in R.~Spigler (ed.), \emph{Applied and Industrial Mathematics}, Mathematics and Its Applications \textbf{56}, 101--109 (Kluwer, 1991), \href{https://doi.org/10.1007/978-94-009-1908-2_9}{doi:10.1007/978-94-009-1908-2\_9}. \bibitem{Tu2008} Y.~Tu, The nonequilibrium mechanism for ultrasensitivity in a biological switch: Sensing by Maxwell's demons, \emph{Proc. Natl. Acad. Sci. USA} \textbf{105}, 11737--11741 (2008), \href{https://doi.org/10.1073/pnas.0804641105}{doi:10.1073/pnas.0804641105}. \bibitem{Dharmadhikari1963} S.~W.~Dharmadhikari, Sufficient Conditions for a Stationary Process to be a Function of a Finite Markov Chain, \emph{Ann. Math. Statist.} \textbf{34}, 1033--1041 (1963), \href{https://doi.org/10.1214/aoms/1177704026}{doi:10.1214/aoms/1177704026}. \bibitem{Heller1965} A.~Heller, On Stochastic Processes Derived From Markov Chains, \emph{Ann. Math. Statist.} \textbf{36}, 1286--1291 (1965), \href{https://doi.org/10.1214/aoms/1177700000}{doi:10.1214/aoms/1177700000}. \bibitem{Klein1978} A.~Klein, A generalization of Markov processes, \emph{Ann. Probab.} \textbf{6}, 128--132 (1978), \href{https://doi.org/10.1214/aop/1176995616}{doi:10.1214/aop/1176995616}. \bibitem[Arveson(1986)]{Arveson1986} W.~Arveson, Markov operators and OS-positive processes, \emph{J. Funct. Anal.} \textbf{66}, 173--234 (1986), \href{https://doi.org/10.1016/0022-1236(86)90071-6}{doi:10.1016/0022-1236(86)90071-6}. \bibitem{JorgensenNeebOlafsson2016} P.~E.~T.~Jorgensen, K.-H.~Neeb and G.~\'Olafsson, Reflection Positive Stochastic Processes Indexed by Lie Groups, \emph{SIGMA} \textbf{12}, 058 (2016), \href{https://doi.org/10.3842/SIGMA.2016.058}{doi:10.3842/SIGMA.2016.058}. \bibitem[Grussler and Damm(2012)]{GrusslerDamm2012} C.~Grussler and T.~Damm, A symmetry approach for balanced truncation of positive linear systems, in \emph{51st IEEE Conference on Decision and Control}, 4308--4313 (2012), \href{https://doi.org/10.1109/CDC.2012.6426442}{doi:10.1109/CDC.2012.6426442}. \bibitem{WagnerTimmer2000} M.~Wagner and J.~Timmer, The Effects of Non-Identifiability on Testing for Detailed Balance in Aggregated Markov Models for Ion-Channel Gating, \emph{Biophys. J.} \textbf{79}, 2918--2924 (2000), \href{https://doi.org/10.1016/S0006-3495(00)76529-5}{doi:10.1016/S0006-3495(00)76529-5}. \bibitem{Chen2011} Y.~Chen, On the reversibility of the observed process of three-state hidden Markov model, \href{https://arxiv.org/abs/1101.5482v1}{arXiv:1101.5482v1} (2011). \bibitem[Wu and Jia(2025)]{WuJia2025} B.~Wu and C.~Jia, Parameter Inference and Nonequilibrium Identification for Markov Networks Based on Coarse-Grained Observations, \emph{Phys. Rev. Lett.} \textbf{134}, 087103 (2025), \href{https://doi.org/10.1103/PhysRevLett.134.087103}{doi:10.1103/PhysRevLett.134.087103}. The explicit binary rotation cited here is in Appendix~A of the author preprint \href{https://arxiv.org/abs/2406.19586v2}{arXiv:2406.19586v2}. \bibitem[Siekmann(2026)]{Siekmann2026} I.~Siekmann, Modelling ion channels with a view towards identifiability, \emph{Bull. Math. Biol.} \textbf{88}, 2 (2026), \href{https://doi.org/10.1007/s11538-025-01558-3}{doi:10.1007/s11538-025-01558-3}. \bibitem{Gu2026} J.~Gu, Excursion-Resolved Thermodynamic Inference without Observing Reverse Transitions, \href{https://arxiv.org/abs/2609.39205v2}{arXiv:2609.39205v2} (2026). \bibitem{Biddle2026} J.~W.~Biddle, A method for calculating entropy production without knowing transition rates or dwell times, \href{https://arxiv.org/abs/2605.27635}{arXiv:2605.27635} (2026). \bibitem{HartichGodec2024} D.~Hartich and A.~Godec, Comment on ``Inferring broken detailed balance in the absence of observable currents'', \emph{Nat. Commun.} \textbf{15}, 8678 (2024), \href{https://doi.org/10.1038/s41467-024-52602-0}{doi:10.1038/s41467-024-52602-0}. \bibitem{BiskerReply2024} G.~Bisker, I.~A.~Mart\'inez, J.~M.~Horowitz and J.~M.~R.~Parrondo, Reply to: Comment on ``Inferring broken detailed balance in the absence of observable currents'', \emph{Nat. Commun.} \textbf{15}, 8679 (2024), \href{https://doi.org/10.1038/s41467-024-52603-z}{doi:10.1038/s41467-024-52603-z}. \bibitem[Fanizza et~al.(2024)]{FanizzaLumbrerasWinter2024} M.~Fanizza, J.~Lumbreras and A.~Winter, \emph{Quantum Theory in Finite Dimension Cannot Explain Every General Process with Finite Memory}, \emph{Commun. Math. Phys.} \textbf{405}, 50 (2024), \href{https://doi.org/10.1007/s00220-023-04913-4}{doi:10.1007/s00220-023-04913-4}. \bibitem{Esposito2012} M.~Esposito, Stochastic thermodynamics under coarse graining, \emph{Phys. Rev. E} \textbf{85}, 041125 (2012), \href{https://doi.org/10.1103/PhysRevE.85.041125}{doi:10.1103/PhysRevE.85.041125}. \bibitem{PolettiniEsposito2017} M.~Polettini and M.~Esposito, Effective thermodynamics for a marginal observer, \emph{Phys. Rev. Lett.} \textbf{119}, 240601 (2017), \href{https://doi.org/10.1103/PhysRevLett.119.240601}{doi:10.1103/PhysRevLett.119.240601}. \bibitem{Ehrich2021} J.~Ehrich, Tightest bound on hidden entropy production from partially observed dynamics, \emph{J. Stat. Mech.} \textbf{2021}, 083214 (2021), \href{https://arxiv.org/abs/2105.08803}{arXiv:2105.08803}. \bibitem{Martinez2019} I.~A.~Mart\'inez, G.~Bisker, J.~M.~Horowitz and J.~M.~R.~Parrondo, Inferring broken detailed balance in the absence of observable currents, \emph{Nat. Commun.} \textbf{10}, 3542 (2019), \href{https://doi.org/10.1038/s41467-019-11051-w}{doi:10.1038/s41467-019-11051-w}. \bibitem{FredkinRice1986} D.~R.~Fredkin and J.~A.~Rice, On aggregated Markov processes, \emph{J. Appl. Probab.} \textbf{23}(1), 208--214 (1986), \href{https://doi.org/10.2307/3214130}{doi:10.2307/3214130}. \bibitem{Kienker1989} P.~Kienker, Equivalence of aggregated Markov models of ion-channel gating, \emph{Proc. R. Soc. Lond. B} \textbf{236}, 269--309 (1989), \href{https://doi.org/10.1098/rspb.1989.0024}{doi:10.1098/rspb.1989.0024}. \bibitem{Fliess1975} M.~Fliess, S\'eries rationnelles positives et processus stochastiques, \emph{Ann. Inst. H. Poincar\'e Sect. B} \textbf{11}, 1--21 (1975), \href{https://www.numdam.org/item/AIHPB_1975__11_1_1_0/}{Numdam original}. \bibitem[Aznagulov(2026)]{Aznagulov2026Architecture} M.~Aznagulov, Thermodynamic Realizability of Hidden Markov Processes: Attainment, Observable Certificates, and the Price of Architecture, \href{https://arxiv.org/abs/2609.14623v1}{arXiv:2609.14623v1} (2026). \bibitem{Combettes2018} P.~L.~Combettes, Monotone operator theory in convex optimization, \emph{Math. Program.} \textbf{170}, 177--206 (2018), \href{https://pcombet.math.ncsu.edu/mp5.pdf}{author manuscript}. \bibitem{LevinPeres2017} D.~A.~Levin and Y.~Peres, \emph{Markov Chains and Mixing Times}, second edition (2017), with contributions by E.~L.~Wilmer, \href{https://pages.uoregon.edu/dlevin/MARKOV/markovmixing.pdf}{author-hosted second edition}. \end{thebibliography} \appendix \section{Reflection reconstruction and the low-rank theorem} \label{app:rank} We prove Theorem~\ref{thm:rank} directly from the observed law. The Hilbert-space reconstruction is standard reflection-positive structure \citep{Klein1978,JorgensenNeebOlafsson2016}; the details below also specify the finite-dimensional and label-compatible contract used here. \subsection{The finite reflected representation} Let $\mathcal C$ be the real span of bounded future cylinder functions. Reversibility makes \eqref{eq:reflection-form} symmetric; reflection positivity makes it a semidefinite form. For a supplied $n$-state CTMC, midpoint conditioning writes every reflected cylinder matrix as a product through $\mathbb R^n$. Consequently the quotient $\mathcal H$ is finite-dimensional, of dimension $r\le n$. Write $[f]$ for a class and $v=[1]$. Then $\|v\|=1$. Time translation $U_t f$ satisfies $[U_t f,g]_R=[f,U_t g]_R$ by stationarity and reversal. This identity shows that a null function remains null after translation. Hence $U_t$ descends to a self-adjoint semigroup $T_t$ on $\mathcal H$. Stochastic continuity of the supplied finite CTMC makes its matrix entries continuous in a cylinder basis, so $T_t=e^{tA}$ with $A=A^{\mathsf T}$ and $Av=0$. For bounded representatives the entries $[f,U_tg]_R$ are uniformly bounded in $t$. In finite dimension this excludes positive eigenvalues, giving $A\le0$. Multiplication by $1_{\{Y_0=a\}}$ is self-adjoint for the reflected form and therefore preserves its null space. It descends to an orthogonal projection $E_a$. These projections are disjoint and sum to the identity. Set $\mathcal H_a=E_a\mathcal H$, $v_a=E_av$ and $p_a=\|v_a\|^2>0$. For positive time increments, the cylinder probabilities are \begin{equation} P(Y_0=a_0,Y_{t_1}=a_1,\ldots, Y_{t_1+\cdots+t_m}=a_m) =v^{\mathsf T}E_{a_0}e^{t_1A}E_{a_1}\cdots e^{t_mA}E_{a_m}v. \label{eq:canonical-cylinder} \end{equation} This follows by representing the future cylinder as the indicated product applied to $v$, and pairing it with $v$. Mixing of the supplied finite irreducible CTMC gives $[f,U_tg]_R\to\E f\E g$. Thus $e^{tA}\to vv^{\mathsf T}$ and $\ker A=\operatorname{span}\{v\}$. Cylinder indicators are dense in ordinary half-path $L^2(P)$. For their differences, Cauchy--Schwarz and reversal give $0\le[f,f]_R\le\|f\|_{L^2(P)}^2$. Therefore measurable finite-window events have classes in the same closed, finite-dimensional quotient. No additional rank arises when classifiers are allowed to be measurable rather than cylindrical. \subsection{The conditional-rank dictionary at every sampling interval} \label{app:rank-dictionary} For $\Delta>0$, let $n_\Delta(a)$ be the rank of the word matrix $[P(s\,a\,t)]$ on the $\Delta$ grid, with arbitrary finite past and future words, including empty words. Reversing the past-word row indexing identifies this matrix with the reflected Gram matrix of the $a$-restricted future grid-word vectors. Hence its rank is their span dimension. Let $W_\Delta$ be the span of all full future grid-word vectors $E_{a_0}T_\Delta E_{a_1}\cdots T_\Delta E_{a_m}v$. It contains $v$, is invariant under every $E_a$, and is invariant under $T_\Delta$: prepending an arbitrary current label and summing over that label gives the shifted word. If $\lambda_1,\ldots,\lambda_q$ are the distinct real eigenvalues of $A$, the numbers $e^{\Delta\lambda_j}$ are distinct. Real polynomial interpolation therefore expresses every $T_t$ as a polynomial in $T_\Delta$. Explicitly, the polynomial \begin{equation} p_t(z)=\sum_{j=1}^q e^{t\lambda_j} \prod_{\ell\ne j}\frac{z-e^{\Delta\lambda_\ell}} {e^{\Delta\lambda_j}-e^{\Delta\lambda_\ell}} \end{equation} satisfies $T_t=p_t(T_\Delta)$. Thus $W_\Delta$ is invariant under all $T_t$ and label masks. These products from $v$ span every future cylinder vector, so $W_\Delta=\mathcal H$. Its $a$-restricted word span is $E_aW_\Delta=\mathcal H_a$, proving $n_\Delta(a)=\dim\mathcal H_a$ for every $\Delta>0$. In particular, at $r\le d+1$, Dharmadhikari's low-conditional-rank corollary already gives an ordinary $r$-state discrete-time realization of each sampled law. The remaining continuous-time detailed-balance obligation is addressed below, rather than by choosing unrelated sampled representations or matrix logarithms. \subsection{Future-event cones and killed semigroups} For each label let $C_a$ be the closed conical hull of the vectors $x_F=[1_F]$ for future cylinder events contained in $\{Y_0=a\}$. These vectors span $\mathcal H_a$, and \begin{equation} \langle x_F,x_G\rangle=P(\theta F\cap G)\ge0. \label{eq:acute} \end{equation} Thus $C_a$ is acute and pointed. It contains $v_a$ and $v_a-x_F$. Choose an event-vector basis $x_1,\ldots,x_k$. For $\sum_i|c_i|\le1$, each vector $v_a+\sum_i c_ix_i$ is a nonnegative combination of $v_a$, $x_i$ and $v_a-x_i$: use $v_a+x_i$ for positive coefficients and $v_a-x_i$ for negative coefficients. Hence $v_a$ is interior to $C_a$. Acuteness and an interior ball imply $\langle v_a,x\rangle\ge\epsilon\|x\|$ on $C_a$ for some $\epsilon>0$: pair $x$ with $v_a-\epsilon x/\|x\|\in C_a$. Put $A_a=E_aAE_a|_{\mathcal H_a}$. The products $(E_ae^{tA/n}E_a)^n$ preserve $C_a$, because on an event vector they require label $a$ at the prefix grid times, followed by the shifted event. They converge on $\mathcal H_a$ to $e^{tA_a}$. Closure gives invariance of $C_a$; self-adjointness gives invariance of its dual $C_a^*$. For distinct labels, the short-time event vectors give \begin{equation} E_aAv_b=\lim_{t\downarrow0}t^{-1}E_ae^{tA}v_b\in C_a. \label{eq:cross-positive} \end{equation} \subsection{An orthogonal positive realization} Since every active label space is nonzero, $r\ge d$. If $r=d$, choose its unit vectors $s_a=v_a/\sqrt{p_a}$. Equation~\eqref{eq:cross-positive} makes every off-diagonal entry of $A$ in this basis nonnegative. If $r=d+1$, exactly one label space, say $\mathcal H_a$, has dimension two. Choose orthonormal coordinates $s=v_a/\sqrt{p_a}$ and $u$. Its normalized primal cone slice is \begin{equation} C_a\cap\{\langle s,x\rangle=1\} =\{s+yu:\ell\le y\le h\},\qquad \ell<00$, with $-LU\ge1$. Write $A_a$ in this basis as $\left(\begin{smallmatrix}\alpha&\gamma\\\gamma&\delta\end{smallmatrix}\right)$. The initial projective derivative under $e^{tA_a}$ is \begin{equation} F(y)=\gamma+(\delta-\alpha)y-\gamma y^2. \end{equation} Dual-cone invariance gives $F(U)\le0$ at its upper endpoint. With $k=\sqrt{1+U^2}$ choose orthonormal rows \begin{equation} j_-=(Us-u)/k,\qquad j_+=(s+Uu)/k. \end{equation} Their slopes $-1/U$ and $U$ both belong to $[L,U]$, so both are in $C_a^*$. Their transformed internal entry is \begin{equation} \langle j_-,A_aj_+\rangle =\frac{U(\alpha-\delta)+\gamma(U^2-1)}{1+U^2} =-\frac{F(U)}{1+U^2}\ge0. \label{eq:rotation-rate} \end{equation} Equation~\eqref{eq:cross-positive} makes their entries against every other unit label vector $s_b$ nonnegative. Entries between the other one-dimensional blocks are nonnegative for the same reason. No sign restriction on $\gamma$ is required. Let $J$ be this block-orthogonal change of basis, using the unit label vectors on all other blocks. Then $B=JAJ^{\mathsf T}$ is symmetric Metzler and $w=Jv>0$. In the two-dimensional block the coordinates are $\sqrt{p_a}U/k$ and $\sqrt{p_a}/k$. With $D_w=\diag(w)$ define \begin{equation} Q=D_w^{-1}BD_w,\qquad \pi_i=w_i^2. \label{eq:rank-generator} \end{equation} Here $Q\one=0$, its off-diagonals are nonnegative, and $\pi_iq_{ij}=w_iB_{ij}w_j=\pi_jq_{ji}$. Thus $Q$ is a conservative detailed-balance row generator. Its graph is connected: each connected component of a symmetric Metzler matrix with positive null vector contributes a null vector, whereas $\ker B$ is one-dimensional. The change of basis preserves each label space. If $P_a$ denotes the corresponding diagonal label mask, then $P_aJ=JE_a$. Conjugating the semigroups in \eqref{eq:rank-generator} gives \begin{equation} \pi P_{a_0}e^{t_1Q}P_{a_1}\cdots e^{t_mQ}P_{a_m}\one =v^{\mathsf T}E_{a_0}e^{t_1A}E_{a_1}\cdots e^{t_mA}E_{a_m}v. \end{equation} These are all the original cylinders, on the original clock. Coordinate cylinders determine the finite-alphabet cadlag law. Finally any $n$-state CTMC realization factors the reflected matrices through $\mathbb R^n$, so $n\ge r$. This proves both exact realization and minimality. \subsection{A nonrenewal example at rank $d+1$} Take four states with uniform stationary masses, labels $(a,a,b,c)$, and symmetric rates $q_{12}=1$, $q_{13}=3$, $q_{14}=1$, $q_{23}=1$, $q_{24}=3$, $q_{34}=2$. The two distinct probe labels give two independent prediction directions; within $a$, $(1,1)$ and $(3,1)$ span its two coordinates. Thus $d=3$ and $r=4$. The killed $a$-block is $\left(\begin{smallmatrix}-5&1\\1&-5\end{smallmatrix}\right)$, whose next-exit-to-$b$ probabilities are $(2/3,1/3)$. Entry from $b$ gives hidden weights $(3/4,1/4)$, whereas entry from $c$ gives $(1/4,3/4)$. The probability of leaving $a$ next for $b$ is consequently $7/12$ or $5/12$, respectively. Entrance-label dependence is compatible with the theorem; binary alternating renewal is not assumed. \section{Joint compression of the binary rank-four profile} \label{app:binary-compression} We prove Theorem~\ref{thm:binary-compression}. The generator acts on column functions. At detailed balance, conditional expectation of a future cylinder $f$ given $X_0$ satisfies $[f,g]_R=\langle v_f,v_g\rangle_\pi$ by midpoint conditioning. Their span contains $\one$ and is invariant under label masks and $e^{tQ}$. Finite dimension permits differentiation at zero, so it is $Q$-invariant. Conversely, the smallest space with these invariances contains every future cylinder vector. Thus that span is $V$, and $V$ is isometric to $\mathcal H$; in particular $r=4$. \paragraph{Features and the complete cross action.} Put $\alpha=\pi(a)>0$, $\beta=\pi(b)>0$ and use conditional stationary measures $\mu_i=\pi_i/\alpha$ on $a$ and $\nu_j=\pi_j/\beta$ on $b$. Choose features $x,y$ supported on their respective labels with conditional mean zero and second moment one. Each $f\in V$ is uniquely $(c_a+ux,c_b+vy)$. Write their extrema as $L_x<00$. All features are dimensionless; the four coefficients, drifts and generator entries have inverse-time units. \paragraph{Joint selection of the two replacement pairs.} Nonnegative original rates give $L_yk_a(x_i)\le m_a(x_i)\le R_yk_a(x_i)$. The two differences are affine in $x$, hence nonnegative on the whole interval between its attained extrema. Consequently \begin{equation} H(x,-1/R_y)=\frac{R_yk_a(x)-m_a(x)}{R_y}\ge0,\qquad H(x,-1/L_y)=\frac{m_a(x)-L_yk_a(x)}{-L_y}\ge0. \label{eq:binary-strip-endpoints} \end{equation} Affinity gives $H\ge0$ on $[L_x,R_x]\times I_y$. Applying the reverse destination-mean bounds in \eqref{eq:binary-reverse-action} gives $H\ge0$ on $I_x\times[L_y,R_y]$. No pointwise escape rate is divided out; zero cross rows are allowed. Bilinear interpolation expresses $H(0,0)=\kappa_0>0$ as a strictly positive convex combination of its four outer-corner values. Choose an outer corner $(x_*,y_*)$ with positive $H$ and replace each feature law by the two points $\{x_*,-1/x_*\}$ and $\{y_*,-1/y_*\}$. The reciprocal points are endpoints of $I_x,I_y$. Thus the other three new product corners lie in the nonnegative strips. Naming the points $x_+>0>x_-$ and $y_+>0>y_-$, give them weights \begin{equation} \begin{gathered} x_+x_-=-1,\qquad \mu'_+=\frac1{1+x_+^2},\quad \mu'_-=\frac{x_+^2}{1+x_+^2},\\ y_+y_-=-1,\qquad \nu'_+=\frac1{1+y_+^2},\quad \nu'_-=\frac{y_+^2}{1+y_+^2}. \end{gathered} \label{eq:binary-replacement-measures} \end{equation} These positive measures match mass, mean and second moment. The retained corner must be selected jointly: independent choices of endpoints would not establish its cross-rate sign. \paragraph{Internal rates and generator intertwining.} Let $T_a=Q_{aa}+\diag(k_a(x_i))$ be the conservative internal generator on $a$. It is reversible for $\mu$ even if its graph is disconnected. Since $(Qx)|_a=h_a(x)$ is affine, \begin{equation} T_ax=D_a(x),\qquad D_a(x)=h_a(x)+k_a(x)x, \qquad \E_\mu D_a=0. \label{eq:binary-internal-drift} \end{equation} This drift has degree at most two, so the replacement measure also has $\E_{\mu'}D_a=0$. At a retained original maximum the CTMC maximum principle gives $D_a(x_*)\le0$; at a retained minimum it gives $D_a(x_*)\ge0$. Positive weights and zero mean drift therefore give $D_a(x_+)\le0\le D_a(x_-)$ in either case. The rates \begin{equation} q'_{a_+,a_-}=-\frac{D_a(x_+)}{x_+-x_-},\qquad q'_{a_-,a_+}=\frac{D_a(x_-)}{x_+-x_-} \label{eq:binary-internal-rates} \end{equation} are nonnegative, realize the drift and satisfy detailed balance for $\mu'$. The identical construction with $D_b(y)=h_b(y)+k_b(y)y$, where $(Qy)|_b=h_b(y)$, gives both internal $b$ rates. Zero drifts remove an internal edge without altering the construction. Set $\pi'_{a_e}=\alpha\mu'_e$, $\pi'_{b_f}=\beta\nu'_f$ for $e,f\in\{+,-\}$, and set the cross rates to \begin{equation} q'_{a_e,b_f}=\nu'_fH(x_e,y_f),\qquad q'_{b_f,a_e}=\frac\alpha\beta\mu'_eH(x_e,y_f). \label{eq:binary-cross-rates} \end{equation} They are nonnegative and have symmetric conductances $\alpha\mu'_e\nu'_fH(x_e,y_f)$. Choosing diagonals as minus outgoing sums defines a conservative detailed-balance generator $Q'$. The first two moments of $\nu'$ give $\sum_fq'_{a_e,b_f}=k_a(x_e)$ and $\sum_fq'_{a_e,b_f}y_f=m_a(x_e)$; the moments of $\mu'$ give both reverse actions in \eqref{eq:binary-reverse-action}. Let $U$ evaluate $f=(c_a+ux,c_b+vy)$ at the four new points. At $a_e$, the action on $Uf$ is \begin{align*} (Q'Uf)_{a_e} &=uD_a(x_e)+(c_b-c_a-ux_e)k_a(x_e)+vm_a(x_e)\\ &=uh_a(x_e)+(c_b-c_a)k_a(x_e)+vm_a(x_e) =(UQf)_{a_e}. \end{align*} At $b_f$ the same calculation gives $vh_b(y_f)+(c_a-c_b)k_b(y_f)+um_b(y_f)=(UQf)_{b_f}$. Thus every row action, including its escape rate, is preserved: \begin{equation} Q'U=U(Q|_V). \label{eq:binary-intertwining} \end{equation} \paragraph{Full law, clock and minimality.} The two distinct points in each label make $U$ a bijection onto all four-state functions. It preserves $\one$ and both label masks. Moment matching also gives, for $g=(\widetilde c_a+\widetilde ux,\widetilde c_b+\widetilde vy)$, \begin{equation} \langle f,g\rangle_\pi =\alpha(c_a\widetilde c_a+u\widetilde u) +\beta(c_b\widetilde c_b+v\widetilde v) =\langle Uf,Ug\rangle_{\pi'}. \label{eq:binary-isometry} \end{equation} Exponentiating \eqref{eq:binary-intertwining} at the same real time and inserting the label masks transports every chronological cylinder function. Pairing with $\one$ through \eqref{eq:binary-isometry} preserves its stationary probability. The finite-alphabet cadlag laws are therefore equal on the original physical clock. Irreducibility of $Q$ gives $\ker(Q|_V)=\operatorname{span}\{\one\}$, so the bijective intertwining gives the same kernel for $Q'$. The detailed-balance Dirichlet form $-\langle f,Q'f\rangle_{\pi'}=\tfrac12\sum_{i,j} \pi'_iq'_{ij}(f_i-f_j)^2$ has one kernel dimension per connected component. Hence $Q'$ is irreducible, including cases with missing internal edges or zero cross corners. The reflected rank is four, and midpoint conditioning bounds it by the state count of any finite CTMC presentation. Four is therefore minimal. \section{A common CP hierarchy with no finite equilibrium realization} \label{app:allcp} We prove Theorem~\ref{thm:allcp} by an explicit rational construction. All displayed numerical rates use one fixed inverse-time unit. \subsection{The core regression and its separating identity} Index five points by $i=0,\ldots,4$ and put \begin{equation} \begin{gathered} p_0=(1,0),\quad p_1=(0,1),\quad p_2=(-1,0),\\ p_3=(-3/5,-4/5),\quad p_4=(3/5,-4/5),\qquad \omega=(5,8,5,5,5)/28. \end{gathered} \label{eq:polygon-data} \end{equation} They have weighted mean zero, covariance $\Sigma=\diag(17/35,18/35)$, and unit norm. Define the directed stationary flows by the following table; unlisted entries are zero: \begin{center} \begin{tabular}{cccccc} \toprule $(i,j)$&$F_{ij}$&$(i,j)$&$F_{ij}$&$(i,j)$&$F_{ij}$\\ \midrule $(0,1)$&$11/126$&$(0,3)$&$5/126$&$(0,4)$&$5/72$\\ $(1,3)$&$25/63$&$(1,4)$&$25/63$&$(2,1)$&$11/126$\\ $(2,3)$&$5/72$&$(2,4)$&$5/126$&$(3,1)$&$13/42$\\ $(3,2)$&$11/56$&$(4,0)$&$11/56$&$(4,1)$&$13/42$\\ \bottomrule \end{tabular} \end{center} Direct rational calculation (reproduced by the ancillary checker \texttt{anc/check\_certificate.py}) gives $\sum_jF_{ij}=\sum_jF_{ji}$ and $\sum_jF_{ij}(p_j-p_i)=\omega_i\diag(-1,-5)p_i$. For $i\ne j$ set \begin{equation} R_{ij}=F_{ij}/\omega_i+\omega_j/10,\qquad R_{ii}=-\sum_{j\ne i}R_{ij}. \label{eq:polygon-R} \end{equation} The refresh term preserves $\omega$ and makes every off-diagonal positive. Thus $R$ is stationary, irreducible and reciprocal, and its coordinate drift is $Lp_i$, where $L=\diag(-\alpha,-\beta)$, $\alpha=11/10$ and $\beta=51/10$. No symmetric nonnegative conductances $C_{ij}=C_{ji}\ge0$ can satisfy \begin{equation} \sum_{j\ne i}C_{ij}(p_j-p_i)=\omega_iLp_i. \label{eq:forbidden-regression} \end{equation} Indeed choose $y_0=y_1=y_4=0$, $y_2=(-35/3,35/3)$ and $y_3=(-21,7)$. For unordered pairs let $d_{ij}=(y_i-y_j)\mathbin\cdot(p_j-p_i)$. All $d_{ij}\le0$; their only nonzero values are \begin{equation} d_{02}=-70/3,\quad d_{03}=d_{24}=-28,\quad d_{34}=-126/5. \end{equation} But $\sum_i\omega_i y_i\mathbin\cdot Lp_i =\beta-13\alpha/3=1/3$. Multiplying \eqref{eq:forbidden-regression} by $y_i$ and summing would give $1/3=\sum_{i0$, so radius saturation forces each $\xi_s$ to be one of the five vertices. Let $S_i$ be the competing core states at vertex $p_i$, and let $\widehat\omega_i=\sum_{s\in S_i}\mu_s$. The $xy$ moment gives $\widehat\omega_3=\widehat\omega_4$; the $x$ mean gives $\widehat\omega_0=\widehat\omega_2$; the $y$ mean gives $\widehat\omega_1=(8/5)\widehat\omega_3$. Normalization and the $x^2$ moment then give $\widehat\omega=\omega$. These are inner-product identities; they do not apply $U$ to quadratic products or assume that it preserves multiplication. The facet sum fixes total exit to all probe labels at each core state to $\lambda$. The original coordinate drift is $(L-\lambda I_2)p_i$; since the coordinates vanish off the core, cancelling that loss in \eqref{eq:polygon-intertwiner} gives \begin{equation} \sum_{t:a}\widehat q_{st}(\xi_t-p_i)=Lp_i, \qquad s\in S_i. \end{equation} Define the aggregated conditional stationary conductances \begin{equation} C_{ij}=2\sum_{s\in S_i,t\in S_j} \widehat\pi_s\widehat q_{st},\qquad i\ne j. \end{equation} They are symmetric and nonnegative by detailed balance. Multiplying the drift identity by $2\widehat\pi_s$ and summing over $S_i$ gives precisely \eqref{eq:forbidden-regression}, contradicting its rational separating certificate. Transitions within a vertex class contribute zero; vertex grouping need not be a Markov lumping. The argument excludes every finite detailed-balance realization of the complete same-clock law, and completes the proof. \section{Half-path factorability and symmetric extensions} \label{app:extensions} In this appendix $W=\mathcal W$ is the observable space of Appendix~\ref{app:allcp}. The extension corollary proves Corollary~\ref{cor:general-extension}. \begin{lemma}[Half-path factorability] \label{lem:halfpath-factorability} Let $Y$ be a stationary reversible finite-alphabet c\`adl\`ag process. On $\mathcal L=L^2(\mathcal F_+,\mathbb P)$, let $\mathsf K=E_+\theta|_{\mathcal L}$ be its reflected half-path operator. Every matched finite-window partition matrix is classically completely positive if and only if $\mathsf K=B^*B$ for an order-preserving bounded operator $B:\mathcal L\to L^2(\nu)$. \end{lemma} \begin{proof} Let $P_n$ be conditional expectation onto increasing finite coordinate fields whose union generates $\mathcal F_+$. Their atoms $A_i$ of positive probability $p_i$ give the positive orthonormal basis $e_i=\mathbf1_{A_i}/\sqrt{p_i}$. In this basis the matrix of $P_n\mathsf K P_n$ is \[ C_{ij}=\frac{\mathbb P(\theta A_i\cap A_j)}{\sqrt{p_i p_j}}. \] Complete positivity is preserved by this positive diagonal rescaling. Writing $C=Z^{\mathsf T}Z$ with $Z\ge0$ gives $P_n\mathsf K P_n=\sum_\ell h_\ell\otimes h_\ell$, where $h_\ell=\sum_i Z_{\ell i}e_i\ge0$. The compressions converge strongly to $\mathsf K$ and have norm at most one, hence also converge in the $\sigma$-weak topology. Arveson's closed-cone characterization of factorability then applies \citep[Theorem~1.2.7]{Arveson1986}. Conversely, if $\mathsf K=B^*B$, put $z_i=B\mathbf1_{A_i}\ge0$. The tested matrix is $M_{ij}=\int z_i z_j\,d\nu$. Normalize by $w=\sum_i z_i$: on $\{w>0\}$, the vector $(z_i/w)_i$ lies in the probability simplex, and $w^2d\nu$ is a probability measure because $\langle\mathbf1,\mathsf K\mathbf1\rangle=1$. Thus $M$ belongs to the compact convex hull of $\{uu^{\mathsf T}:u\ge0,\ \sum_i u_i=1\}$. A finite-dimensional convex decomposition supplies a finite nonnegative Gram factor. \end{proof} \begin{lemma}[Finite-valued reversible reduction] \label{lem:finite-valued-reversible-reduction} Let $(T_t)_{t\ge0}$ be a self-adjoint Markov contraction semigroup on $L^2(\mu)$, with $\mu$ a probability measure. Let $V$ be a finite-dimensional real invariant subspace containing $\mathbf1$ and invariant under every deterministic label multiplier. Assume continuity of $T_t|_V$ at zero, and write $T_t|_V=e^{tA_V}$. Suppose that its functions are constant on a finite measurable partition. Then the observed law has an exact finite detailed-balance CTMC realization at the same clock. \end{lemma} \begin{proof} Remove null classes, and merge classes with identical joint values of a basis $h=(h_1,\ldots,h_k)$ of $V$. Write these distinct value vectors as $v_i$ and their positive masses as $\mu_i$. The finite-dimensional restriction gives $(T_t h-h)/t\to A_Vh$ in $L^2$ for every $h\in V$. For $i\ne j$, set \[ c_{ij}(t)=\frac{\langle\mathbf1_{S_i},T_t\mathbf1_{S_j}\rangle}{t}. \] These quantities are symmetric and nonnegative. Since $T_t\mathbf1=\mathbf1$, an exact expansion gives \[ E_t:=\sum_{\ell=1}^k \frac{\langle h_\ell,(I-T_t)h_\ell\rangle}{t} =\frac12\sum_{i\ne j}c_{ij}(t)|v_i-v_j|^2 \longrightarrow-\sum_\ell\langle h_\ell,A_Vh_\ell\rangle. \] If there is only one class, the law is constant and its one-state realization is immediate. Otherwise the minimum separation between the $v_i$ is positive. The boundedness of $E_t$ therefore bounds every $c_{ij}(t)$. A common subsequence $t_n\downarrow0$ yields finite, symmetric, nonnegative limits $c_{ij}$. Moreover, \[ \sum_{j\ne i}c_{ij}(t)(v_j-v_i) =\int_{S_i}\frac{T_t h-h}{t}\,d\mu \longrightarrow\int_{S_i}A_Vh\,d\mu. \] Define $q_{ij}=c_{ij}/\mu_i$ for $i\ne j$ and choose the diagonal for zero row sums. Then $Q$ is a finite conservative row generator and $\mu_iq_{ij}=\mu_jq_{ji}$. Since $A_Vh\in V$ is class-constant, the last identity says $Qh=A_Vh$ on each class. Consequently $e^{tQ}$ and $T_t$ agree on $V$. The labels are constant on the classes, and label-mask invariance keeps all chronological cylinder products starting from $\mathbf1$ in $V$. Their stationary probabilities therefore agree in both realizations. This proves equality of the full observed law. The partition indicators were used only as $L^2$ test functions; no generator-domain condition on them, hidden finite jump rate, or strong lumpability was required. \end{proof} \begin{corollary}[Exclusion of arbitrary-state equilibrium extensions] \label{cor:arbitrary-state-equilibrium-exclusion} The rational ten-state, six-label construction has no exact realization as an instantaneous deterministic Borel observation of a stationary two-sided ordinary symmetric Markov process on any standard Borel state space at the same continuous clock. \end{corollary} \begin{proof} Suppose such a realization exists. Its conditional future vectors and those of the original model have the same reflected Gram kernel. Their correspondence therefore gives an isometry $U$ from the original eight-dimensional observable space $W$ into $L^2(\mu)$. Its transition operators $T_t$ are self-adjoint Markov contractions. The isometry preserves $\mathbf1$, intertwines the label masks, and satisfies \[ T_tU=Ue^{tQ_0}|_W. \] Thus $V=U(W)$ is an invariant finite-dimensional space with the explicit analytic restriction $T_t|_V=e^{tA_V}$, where $A_V=U(Q_0|_W)U^{-1}$. On the core label $a$, set $\xi=(Ux,Uy)$. For each probe label $b_\ell$, positivity of the transition semigroup gives \[ \mathbf1_a A_V\mathbf1_{b_\ell} =\lim_{t\downarrow0}\frac{\mathbf1_aT_t\mathbf1_{b_\ell}}t =\mathbf1_a g_\ell(\xi)\ge0 \qquad\text{in }L^2. \] Hence $\xi$ lies almost surely in the polygon defined by the five probe inequalities. Isometry fixes its conditional mean to zero and its conditional covariance to $\operatorname{diag}(17/35,18/35)$, so $\mathbb E[|\xi|^2\mid a]=1$. The polygon is contained in the unit disk and meets its boundary only at the five specified vertices. Saturation therefore forces $\xi$ to take those vertex values almost surely; the moments fix their positive conditional weights to $(5,8,5,5,5)/28$. Off the core label, $Ux=Uy=0$, and each probe is distinguished by its label indicator. Consequently every function in $U(W)$ is constant on the resulting ten-class measurable partition. Lemma~\ref{lem:finite-valued-reversible-reduction} supplies an exact finite detailed-balance CTMC realization, contradicting the rational symmetric-conductance obstruction of the construction. \end{proof} \begin{remark} The construction has the explicit order-preserving factor $Bf(s)=\mathbb E[f(Y_{[0,\infty)})\mid X_0=s]$ into $L^2(\pi_0)$, and $\mathsf K=B^*B$. Its obstruction is therefore distinct from the nonfactorability mechanism used by Arveson. The corollary does not by itself imply a positive entropy-production infimum; the separate closure argument in Appendix~\ref{app:closure} establishes that stronger conclusion for this law. \end{remark} \section{An explicit finite-observation separation certificate} \label{app:closure} We prove Theorem~\ref{thm:closure} for the unchanged generator \eqref{eq:polygon-generator}. A rational polynomial in its known generator defines a finite-lag prediction action. Mismatch with that action is measured by observed residuals; it is never assumed away. No competing generator is differentiated, reconstructed by a logarithm or subject to a rate bound. \subsection{Rational features and a finite-lag observable action} Fix \begin{equation} \begin{gathered} h=10^{-3},\qquad t=10^{-7},\qquad \rho=0,\qquad N=12,\\ \beta_t=\frac{25}{2352t}+\frac{25}{32},\qquad a_t=\frac{275}{9408t}+\frac{225}{256},\\ C_\delta=63843,\qquad C_G=\left(\frac{23}h+36\right)a_t +\left(12+\frac{12}h\right)\beta_t,\qquad C_R=\frac{491}{48h}+\frac{12960}{h^2}. \end{gathered} \label{eq:certificate-parameters} \end{equation} For a future observed record, set $e=1_{\{Y_0=a\}}$, $e_\ell=1_{\{Y_0=b_\ell\}}$ and \begin{equation} \begin{gathered} F_\ell=\frac{1_{\{Y_0=a,Y_t=b_\ell\}}}{t},\qquad F_x=\frac{F_2-F_1}{4},\qquad F_y=e-\frac{F_1+F_2}{4},\\ F_{r,\ell}=F_\ell-b_\ell e-m_{\ell,x}F_x-m_{\ell,y}F_y \quad(\ell=3,4,5). \end{gathered} \label{eq:certificate-features} \end{equation} Here $b_\ell,m_\ell$ are the fixed facet coefficients in \eqref{eq:polygon-facets}, not random probe indicators. Let $\phi=(e,x_0,y_0,e_1,\ldots,e_5)$ be the ideal target basis, where $x_0,y_0$ take the original vertex coordinates on core states and vanish on probes. Define $C$ by $Q_0\phi_j=\sum_iC_{ij}\phi_i$. Equation~\eqref{eq:polygon-action} specifies it explicitly: \begin{equation} C=\begin{pmatrix} -\lambda&0&0&b^{\mathsf T}\\ 0&-(\lambda+\alpha)&0&m_x^{\mathsf T}\\ 0&0&-(\lambda+\beta)&m_y^{\mathsf T}\\ 5b&5\Sigma_{xx}m_x&5\Sigma_{yy}m_y&-5\diag(b) \end{pmatrix},\qquad A_h=\sum_{n=0}^{N}\frac{(hC)^n}{n!}. \label{eq:certificate-finite-action} \end{equation} Here $b=(b_\ell)_{\ell=1}^5$, $m_x=(m_{\ell,x})_{\ell=1}^5$ and $m_y=(m_{\ell,y})_{\ell=1}^5$ are column vectors. The first three basis indices are $e,x,y$, followed by the probes. For the corresponding observed basis $\mathbf F=(e,F_x,F_y,e_1,\ldots,e_5)$, define \begin{equation} F_{Aj}=\sum_i(A_h)_{ij}F_i,\qquad F_{Rj}=\operatorname{shift}_hF_j-F_{Aj} \quad(j=e,x,y),\qquad F_e=e. \label{eq:certificate-euler} \end{equation} Here the shift reads the same functional starting at time $h$. These are finite-action residuals, not generator residuals. On a core coordinate $\xi=(x,y)$ let $A_j(\xi)=(A_h)_{ej}+(A_h)_{xj}x+(A_h)_{yj}y$ and put \begin{equation} H_h(\xi)=(A_x(\xi),A_y(\xi))-\xi A_e(\xi) =c_h+B_h\xi-\xi(k_h\cdot\xi). \label{eq:certificate-lag-drift} \end{equation} Thus $B_h$ is the core coordinate coefficient block minus $(A_h)_{ee}I$ and $k_h=((A_h)_{xe},(A_h)_{ye})$. For $w_i=y_i$ set $z_i=w_i\cdot H_h(p_i)$. The rational polynomial $q_h=d_0+d_1x+d_2y+d_3x^2+d_4xy$ is specified by \begin{equation} \begin{gathered} d_0=\frac{25}{72}(z_3+z_4)+\frac59z_1-\frac18(z_0+z_2), \qquad d_1=\frac{z_0-z_2}{2},\\ d_2=z_1-d_0,\qquad d_3=\frac{z_0+z_2}{2}-d_0,\qquad d_4=\frac54d_1+\frac{25}{24}(z_3-z_4). \end{gathered} \label{eq:certificate-interpolant} \end{equation} This is the inverse vertex-evaluation rule, so $q_h(p_i)=z_i$. Its constant term is core restricted. With reflection $\theta$ read outward from the midpoint, define \begin{align} G_3&=F_{r,3}+\frac{147}{5}te,\qquad G_5=F_{r,5}+\frac{147}{5}te,\nonumber\\ G_4&=F_{r,4}+\frac{842}{25}te-2tF_y,\qquad G_{\mathrm{obs}}=\sum_{\ell=3}^5(\theta G_\ell)G_\ell. \label{eq:certificate-signed-facets} \end{align} \begin{align} q_{\mathrm{obs}}&=d_0e+d_1F_x+d_2F_y +d_3(\theta F_x)F_x+d_4(\theta F_x)F_y,\nonumber\\ \delta_{\mathrm{obs}}&=e-(\theta F_x)F_x-(\theta F_y)F_y,\qquad R_{\mathrm{obs}}=\sum_{j=e,x,y}(\theta F_{Rj})F_{Rj},\nonumber\\ L_{\mathrm{obs}}&=\frac{q_{\mathrm{obs}}}{h} -C_\delta\delta_{\mathrm{obs}}-C_GG_{\mathrm{obs}} -C_RR_{\mathrm{obs}}. \label{eq:certificate-score} \end{align} Only times $-h-t,-h,-t,0,t,h,h+t$ are used. Translation by $h+t$ gives the seven-time tuple in Theorem~\ref{thm:closure}. Every coefficient is rational. \subsection{The universal equilibrium inequality} Here $\E_{\mathrm{core}}$ denotes an unnormalized core expectation. Take any stationary finite detailed-balance competitor, with no rate, dimension or topology restriction. Write $\xi=(x,y)$ for its conditional coordinate features at the hidden midpoint, zero off the core, and $f_\ell,r_\ell$ for its conditional probe and facet residual features. Let $T_h$ be its genuine semigroup and $\bar\phi=\E[\mathbf F\mid X_0]$. Retain the actual residuals \begin{equation} R_j=T_h\bar\phi_j-\sum_i(A_h)_{ij}\bar\phi_i \quad(j=e,x,y). \label{eq:certificate-competitor-residual} \end{equation} As before, $g_1(\xi)=f_1\ge0$, $g_2(\xi)=f_2\ge0$, and $g_\ell(\xi)=f_\ell-r_\ell\ge-r_\ell$ for $\ell=3,4,5$. Set $\eta_f=\max(0,r_3,r_4,r_5)$ and $a=1+5\eta_f/8$. Every facet intercept is at least $8/5$, so $z=\xi/a\in K$. Choose a nearest vertex $v(\xi)$ with a fixed tie rule, and set $v=0$ off the core. Since all vertices have unit norm, this maximizes $z\cdot p_i$. A barycentric representation of $z$ gives $\max_i z\cdot p_i\ge|z|^2$. Consequently \begin{equation} |\xi-v|^2\le1-|\xi|^2+ \frac54\eta_f+\frac{25}{32}\eta_f^2. \label{eq:certificate-rounding-pointwise} \end{equation} Indeed the left side is at most $1+a^2|z|^2-2a|z|^2 =1-|\xi|^2+2a(a-1)|z|^2$, and $|z|\le1$. A first-moment estimate is sharper here. For any distinct unit vertices with minimum distance $d_*$ and any $z$ in their convex hull, take independent barycentric vertices $P_1,P_2$ with mean $z$. Then \[ 1-|z|^2=\tfrac12\E|P_1-P_2|^2 \ge\tfrac{d_*}2\E|P_1-P_2| \ge\tfrac{d_*}2\E|P_1-z| \ge\tfrac{d_*}2\operatorname{dist}(z,V). \] The middle inequality is conditional Jensen; the coefficient $2/d_*$ is attained by two vertices at their midpoint. Here $d_*^2=4/5$, so choose $\kappa=9/4\ge\sqrt5$ before rescaling $z$. It follows that \[ |\xi-v|\le\kappa(1-|\xi|^2) +\frac{55}{16}\eta_f+\frac{225}{256}\eta_f^2. \] Indeed $|\xi-v|\le(a-1)|z|+\kappa(1-|z|^2)$ and $1-|z|^2=1-|\xi|^2+(a^2-1)|z|^2$. Only the combined right side is required to be nonnegative. Let $\gamma_\ell=\E[G_\ell\mid X_0]$. On the core these are $r_\ell+c_\ell$, where $c_3=c_5=147t/5$ and $c_4=t(842/25-2y)$. Nonnegativity of $f_1,f_2$ gives $y=1-(f_1+f_2)/4\le1$, so $c_4\ge792t/25>c_{\min}:=147t/5$. For every real $r$ and $c\ge c_{\min}>0$, \[ r_+\le\frac{(r+c)^2}{4c_{\min}},\qquad r_+^2\le(r+c)^2. \] These inequalities follow from $(r-c)^2\ge0$ when $r>0$; both left sides vanish when $r\le0$. Thus, with \begin{equation} \delta=\E e-\E|\xi|^2,\quad A_G=\E\sum_{\ell=3}^5\gamma_\ell^2,\quad D^2=\E|\xi-v|^2,\quad U=\E|\xi-v|,\quad \mathcal R=\E\sum_{j=e,x,y}R_j^2, \label{eq:certificate-moments} \end{equation} averaging \eqref{eq:certificate-rounding-pointwise} gives \begin{equation} D^2\le S:=\delta+\beta_t A_G,\qquad U\le\frac94\delta+a_t A_G. \label{eq:certificate-rounding} \end{equation} Here $S\ge0$ even when $\delta<0$. Detailed balance and conditional independence identify $\delta$, $A_G$ and $\mathcal R$, and hence $S$, with their reflected observed expressions; $D^2$ remains a realization-dependent rounding error. The $R_j$ remain the competitor's actual finite-action residuals in \eqref{eq:certificate-competitor-residual}. We record two stability estimates. Strong monotonicity and its regularization are classical \citep[author manuscript, pp.~1--2]{Combettes2018}. The certificate below instead uses the first moment and only weak monotonicity; no global field extension is required. \begin{lemma}[Strongly monotone dual-field stability] \label{lem:strict-rounding} Let $V=\{p_1,\ldots,p_N\}\subset\mathbb R^d$ consist of distinct vertices, $N\ge2$. Let $w_1,\ldots,w_N\in\mathbb R^d$ satisfy \[ A_{ij}:=(w_i-w_j)\cdot(p_i-p_j)>0\qquad(i\ne j), \qquad K:=\max_{i0$. Let $S\ge D^2$ and set \[ \Psi_h=\langle\xi,(I-T_h)\xi\rangle_\pi, \qquad p_h=\Pp(e_0=1,e_h=0),\qquad \Lambda_h(S)=\pos{\pos{\sqrt{\Psi_h}-\sqrt S}^{2}-p_h}. \] For $0\le\vartheta<1$, the work in \eqref{eq:strict-rounding-continuous} satisfies \begin{equation} \E[w(e_0,v_0)\cdot(\xi_h-\xi_0e_h)] \le\frac{KS}{1-\vartheta}-\vartheta\rho\Lambda_h(S). \label{eq:observed-transition-penalty} \end{equation} For every $b>0$ it also satisfies the affine bound \begin{equation} \E[w(e_0,v_0)\cdot(\xi_h-\xi_0e_h)] \le-\vartheta\rho\left(\frac{\Psi_h}{1+b}-p_h\right) +\left(\frac K{1-\vartheta}+\frac{\vartheta\rho}{b}\right)S. \label{eq:observed-transition-affine} \end{equation} For conditional features of the observed probes used here, $\Psi_h$, $p_h$ and the certified $S$ are determined by the existing seven-time law. \end{corollary} \begin{proof} Retaining a fraction $\vartheta$ of the negative term in the completed square of Lemma~\ref{lem:strict-rounding} gives \[ W_h\le-\frac\vartheta2G_h+\frac{KD^2}{1-\vartheta},\qquad G_h=\E_{\mathrm{core,core}} A_{v_0,v_h}. \] Put $E_v=\langle v,(I-T_h)v\rangle_\pi$. The unit vertex norms and stationarity give \[ 2(E_v-p_h)=\E_{\mathrm{core,core}}|v_h-v_0|^2\ge0, \qquad G_h\ge2\rho(E_v-p_h). \] The seminorm induced by $I-T_h$ obeys the triangle inequality. Since $0\le I-T_h\le I$, the error $u=\xi-v$ has energy at most $D^2\le S$. Thus $\sqrt{E_v}\ge\pos{\sqrt{\Psi_h}-\sqrt S}$, proving \eqref{eq:observed-transition-penalty}. Alternatively, Young's inequality gives $\Psi_h\le(1+b)E_v+(1+1/b)S$ and proves \eqref{eq:observed-transition-affine}. Finally, reflection and midpoint conditioning identify \[ \Psi_h=\sum_{j=x,y}\E[(\theta F_j)F_j -(\theta F_j)(\operatorname{shift}_hF_j)], \qquad p_h=\Pp(Y_0=a,Y_h\ne a). \] These expressions use a subset of the existing observation times. The work $W_h$ remains an internal rounded-feature term; a strengthened bounded score still requires the interpolation and residual estimates. \end{proof} For the fixed weak field $w_i=y_i$, all $A_{ij}\ge0$ and $M^2\le490<23^2$. Write $H_h(\xi)=c_h+B_h\xi-\xi(k_h\cdot\xi)$ and $M_q=\left(\begin{smallmatrix}d_3&d_4/2\\d_4/2&0\end{smallmatrix}\right)$. For $\xi=p_i+u$ define \[ \ell_i=\nabla q_h(p_i)-B_h^{\mathsf T}w_i +(w_i\cdot p_i)k_h+(k_h\cdot p_i)w_i,\qquad M_i=M_q+\frac{w_i k_h^{\mathsf T}+k_h w_i^{\mathsf T}}2. \] Interpolation gives the exact cancellation \[ q_h(\xi)-w_i\cdot H_h(\xi) =\ell_i\cdot u+u^{\mathsf T}M_i u. \] The displayed rational action and interpolant yield \begin{equation} \begin{gathered} \max_i|\ell_i|^2<(36h)^2,\qquad \max_i\lambda_{\max}(M_i)<12h,\\ \max_i|w_i|^2<491,\qquad \max_i\bigl(|w_i|^2+(w_i\cdot p_i)^2\bigr)<540,\\ \sum_i|(A_h)_{ij}|\le2\ (j=e,x,y),\qquad \sum_{r=0}^4|d_r|\le60h. \end{gathered} \label{eq:certificate-coefficient-bounds} \end{equation} The eigenvalue bound follows from rational Gershgorin upper bounds. Expanding the residual pairing at $p_i$ bounds it by $\sqrt{540}\sqrt{\mathcal R}+\sqrt{491}D\sqrt{\mathcal R}$. Use $\sqrt{491}D\sqrt{\mathcal R} \le12D^2+(491/48)\mathcal R$. The residual identity is \begin{equation} T_h\xi-\xi T_he=H_h(\xi)+R_{(x,y)}-\xi R_e \quad\hbox{on the core}, \label{eq:certificate-probe-cancellation} \end{equation} Combining it with \eqref{eq:weak-first-moment-work} gives \begin{equation} \E_{\mathrm{core}}\frac{q_h(\xi)}h \le\left(\frac{23}h+36\right)U +\left(12+\frac{12}h\right)D^2 +\frac{\sqrt{540}}h\sqrt{\mathcal R} +\frac{491}{48h}\mathcal R. \label{eq:certificate-euler-dual} \end{equation} Substitute both bounds in \eqref{eq:certificate-rounding} with their separate $\delta$ and $A_G$ coefficients; $\delta$ need not be nonnegative. The remaining completed square is \[ \frac{\sqrt{540}}h\sqrt{\mathcal R} \le\frac1{96}+\frac{12960}{h^2}\mathcal R. \] The coefficients in \eqref{eq:certificate-parameters} prove \begin{equation} \E_{P'}L_{\mathrm{obs}}\le\frac1{96} \qquad(P'\in\mathcal E_{\mathrm{fin}}). \label{eq:certificate-equilibrium-bound} \end{equation} There is no square-root rounding budget or inverse-coercivity factor. \subsection{Fixed-target evaluation and bounded normalization} The target generator satisfies $\|Q_0\|_\infty<26$. For $s=t,h$ put $P_s=\sum_{n=0}^{12}(sQ_0)^n/n!$ and \begin{equation} u_s=\frac{2(26s)^{13}}{13!},\qquad \|e^{sQ_0}-P_s\|_\infty\le u_s. \label{eq:certificate-target-remainder} \end{equation} The exponential-series bound uses $26s<1/2$ and concerns only the fixed target. Build the polynomial conditional coordinates $x_i^{\mathrm p},y_i^{\mathrm p}$ from $P_t$ by the same probe formula as the observed features. Their coordinate errors are at most $e_t=u_t/(2t)$ and their magnitudes are below two. Let $\delta^{\mathrm p},q^{\mathrm p},A_G^{\mathrm p}$ denote the corresponding rational stationary sums. Then \begin{align*} \delta&\le\delta^{\mathrm p}+ \sum_{i:a}\pi_i\{2e_t(|x_i^{\mathrm p}|+|y_i^{\mathrm p}|)+2e_t^2\},\\ \E_P(q_{\mathrm{obs}}/h)&\ge q^{\mathrm p}/h- \frac1h\sum_{i:a}\pi_i\bigl\{(|d_1|+|d_2|)e_t +|d_3|(2|x_i^{\mathrm p}|e_t+e_t^2)\\ &\hspace{6em}+|d_4|[(|x_i^{\mathrm p}|+|y_i^{\mathrm p}|)e_t+e_t^2]\bigr\}. \end{align*} The errors in the shifted conditional facets are at most $(5e_t,(4+2t)e_t,5e_t)$. If $\bar\phi^{\mathrm p}$ is the polynomial conditional basis, set $R^{\mathrm p}=P_h\bar\phi^{\mathrm p}-\bar\phi^{\mathrm p}A_h$ in its first three columns. Sup-norm contraction and \eqref{eq:certificate-coefficient-bounds} give $|R_j-R_j^{\mathrm p}|\le3e_t+2u_h$. Bounding each squared component by its rational interval endpoint, the fixed parameters give \begin{equation} \begin{gathered} \E_P(q_{\mathrm{obs}}/h)>0.1134506396,\qquad \delta<1.137856\times10^{-6},\\ A_G<5.51524\times10^{-24},\qquad \mathcal R<8.21447\times10^{-17}. \end{gathered} \label{eq:certificate-target-moments} \end{equation} All four bounds are outward-rounded rational endpoints. The ancillary file \texttt{anc/check\_certificate.py} reproduces them independently: it encloses $e^{sQ_0}$, $s\in\{t,h\}$, by its exact degree-$16$ Taylor polynomial plus an entrywise remainder interval, propagates outward-rounded rational intervals, and evaluates the four moments both through \eqref{eq:polygon-gram} and directly with the stationary reverse chain of $Q_0$ for the reflected half. It also recomputes \eqref{eq:certificate-coefficient-bounds}, the constants in \eqref{eq:certificate-parameters}, and $L_{\mathrm{obs}}$ on all $6^7$ words, confirming $|L_{\mathrm{obs}}|\le B$. The target restricted-adjoint Gram identity \eqref{eq:polygon-gram} identifies these conditional quadratic moments with their reflected observable expectations; microscopic symmetry of $Q_0$ is not assumed. Substitution yields \begin{equation} \E_P L_{\mathrm{obs}}>\frac1{25}. \label{eq:certificate-target-bound} \end{equation} This target calculation is an interval enclosure of the complete finite-lag action, rather than extrapolation of small-$t$ values. The normalization holds on every word over the six labels. The first three observed features have possible values \[ V_t=\left\{(0,0,0),(1,0,1), \left(1,-\frac1{4t},1-\frac1{4t}\right), \left(1,\frac1{4t},1-\frac1{4t}\right)\right\}. \] Every entry of the complete observed basis has magnitude at most $1/(4t)$, and the diameter of $V_t$ is at most $1/(2t)$. Since $\|C\|_1=288/5$, put $r=h\|C\|_1=36/625$; then \[ \|A_h-I\|_1<\frac r{1-r}=\frac{36}{589}<\frac1{10}. \] Write $F_R=(\operatorname{shift}_h\mathbf F_{e,x,y}-\mathbf F_{e,x,y}) -((A_h-I)^{\mathsf T}\mathbf F)_{e,x,y}$. The first difference has norm at most $1/(2t)$, and each entry of the second is at most $\|A_h-I\|_1/(4t)<1/(40t)$. Thus $\|F_R\|\le11/(20t)$ and, by Cauchy--Schwarz, $|R_{\mathrm{obs}}|\le121/(400t^2)$. The rescaled unshifted facet residual vector is $z_j-tc$, where $c=(3,18/5,3)$ and \[ z_j\in\{(0,0,0),(3/4,1/2,-1/4),(-1/4,1/2,3/4), (1,0,0),(0,1,0),(0,0,1)\}. \] Hence $\|F_r\|\le(1+6t)/t$. The shifts contribute at most $93t+1/2$ in Euclidean norm, using $|F_y|\le1/(4t)$. Since $t\le10^{-6}$, this gives $\|G\|\le(1+7t)/t$ and $|G_{\mathrm{obs}}|\le(1+7t)^2/t^2$. Also $|q_{\mathrm{obs}}/h|\le60/t^2$ and $|\delta_{\mathrm{obs}}|\le3/t^2$. These are pointwise word bounds, independent of any generator. Define the rational constants \begin{equation} \begin{gathered} B=\frac{60+3C_\delta+(1+7t)^2C_G+\frac{121}{400}C_R}{t^2},\qquad W=\frac{L_{\mathrm{obs}}+B}{2B},\\ c=\frac12+\frac1{192B},\qquad\eta=\frac{71}{4800B}. \end{gathered} \label{eq:certificate-normalization} \end{equation} They satisfy $0\le W\le1$. The raw gap $1/25-1/96=71/2400$ proves the claimed separation, with $\eta\simeq1.2398\times10^{-26}$. These constants still do not establish practical statistical power. \subsection{Cylinder separation and entropy production} Let $m_0$ index the displayed seven-time tuple, with the current value of $t$, in the enumeration of \eqref{eq:cylinder-metric-main}. Changing the probe lag also changes this tuple; no comparison of its enumeration weight with that of another tuple is required. A $[0,1]$ score gap is bounded by supremum-event total variation, so \begin{equation} \tv(P_{\mathcal T},P'_{\mathcal T})\ge\eta,\qquad D(P,P')\ge2^{-m_0}\eta \quad(P'\in\mathcal E_{\mathrm{fin}}). \label{eq:positive-cylinder-gap} \end{equation} The metric concerns rational coordinate cylinders, rather than uniform path-window total variation. \subsection{Reversibilization and the explicit dissipation floor} For any exact compatible finite irreducible reciprocal realization $Q,\pi,h$, let $Q^\dagger$ be its stationary reverse and set $\overline Q=(Q+Q^\dagger)/2$. Stationarity gives $q^\dagger_{ii}=q_{ii}$, so the two generators have the same escape rate at every state. The comparison $\overline Q$ is irreducible and detailed balanced at $\pi$, with unchanged labels and clock. For a supported unordered pair write $a=\pi_iq_{ij}>0$, $b=\pi_jq_{ji}>0$. The stationary path likelihoods have equal initial and holding-time factors, giving the exact finite-interval identity \begin{equation} \KL(\mathbb P_{Q,[0,T]}\Vert\mathbb P_{\overline Q,[0,T]}) =T\sum_{i0$. For every linear map $U:\mathbb R^m\to\mathcal H$ and $0<\alpha\le1$, \begin{equation} \|T^\alpha U-UA^\alpha\|_{\mathrm{HS}} \le a_*^{\alpha-1}\|TU-UA\|_{\mathrm{HS}}. \label{eq:fractional-prediction-transfer} \end{equation} The coefficient is sharp. Neither invariance of the range of $U$ nor an isometry assumption on $U$ is required. \end{lemma} \begin{proof} Choose an orthonormal eigenbasis $v_j$ of $A$, with eigenvalues $a_j$. For $a\in[a_*,1]$, define \[ g_a(x)=\frac{x^\alpha-a^\alpha}{x-a}\quad(x\ne a), \qquad g_a(a)=\alpha a^{\alpha-1}. \] Concavity gives $0\le g_a(x)\le a^{\alpha-1}$ for $0\le x\le1$: the largest secant slope at fixed $a$ is its value at $x=0$. Spectral calculus therefore gives \[ (T^\alpha-a_j^\alpha I)Uv_j =g_{a_j}(T)(T-a_jI)Uv_j, \qquad \|g_{a_j}(T)\|\le a_*^{\alpha-1}. \] Summing the squared column bounds proves \eqref{eq:fractional-prediction-transfer}, including $\alpha=1$. Taking $T=0$ and $A=a_*I$ gives equality. The same coefficient is approached on the centered coordinate of a reversible two-state chain with equal stationary masses and jump rate $r$, whose time-$h$ eigenvalue is $e^{-2rh}$. Thus unrestricted competing rates prevent a uniform additional factor proportional to $\alpha$ as $\alpha\downarrow0$. This sharpness example concerns the operator inequality, rather than compatibility with the fixed target law. \end{proof} To apply the lemma, retain $h=10^{-3}$ and the known $C,A_h$ from \eqref{eq:certificate-finite-action}. Its reference Gram matrix and symmetric conjugates are \[ G=\diag(1/2,17/70,18/70,1/10,\ldots,1/10),\qquad B_0=G^{1/2}CG^{-1/2},\qquad A=G^{1/2}A_hG^{-1/2}. \] The restricted-adjoint identity makes $B_0$ symmetric. Stationarity of the target gives $\langle f,Q_0f\rangle_\pi\le0$, so its spectrum lies in $[-26,0]$ by $\|Q_0\|_\infty<26$. Consequently $A$ has spectrum in $[a_*,1]$, where $a_*=487/500$: for $0\le x\le26h$, the alternating polynomial $P_{12}(-x)$ lies between $1-x$ and $1$. Its remainder gives \[ \|A-e^{hB_0}\|\le u_h,\qquad u_h=\frac{2(26h)^{13}}{13!}. \] These are bounds on the known target action alone. Take any finite stationary detailed-balance competitor, allowing arbitrary hidden state count, rates, topology and reducibility. Let $\mathbf F=(e,F_x,F_y,e_1,\ldots,e_5)$ be bounded observed functionals of the times $0,t$, where $e,e_\ell$ are the actual label indicators at time zero and $F_x,F_y$ vanish off the core label. The raw features in \eqref{eq:certificate-features}, or specified bounded affine corrections with the same support, are permitted. Define $\bar\phi_j=\E[ F_j\mid X_0]$, and let $\Phi:\mathbb R^8\to L^2(\pi)$ have columns $\bar\phi_j$. Set \begin{equation} \begin{gathered} U=\Phi G^{-1/2},\qquad R_j=T_h\bar\phi_j-\sum_i(A_h)_{ij}\bar\phi_i \quad(j=e,x,y,e_1,\ldots,e_5),\\ \mathcal R_G=\sum_j\frac{\|R_j\|_\pi^2}{G_{jj}},\qquad \mathcal M_G=\sum_j\frac{\|\bar\phi_j\|_\pi^2}{G_{jj}}. \end{gathered} \label{eq:fractional-weighted-residuals} \end{equation} The norms are thus $\|T_hU-UA\|_{\mathrm{HS}}^2=\mathcal R_G$ and $\|U\|_{\mathrm{HS}}^2=\mathcal M_G$. All eight columns are used; the competitor's actual Gram matrix need not equal $G$. \begin{proposition}[Observed control of every facet violation] \label{prop:fractional-observed-facets} For every $00, \end{equation} since the numerator vanishes at zero and has derivative $2\theta^2/(1-\theta^2)^2>0$. The endpoint at $u=1$ is handled by continuity. Jensen gives $J\ge4h(H)=4\ff(1-H^2)$. The positive power series in \eqref{eq:F} proves monotonicity and convexity of $\ff$. For $p=((1+r)/2,(1-r)/2)$ and $q$ with its entries exchanged, $1-H^2=r^2$ and equality holds. Thus this \emph{probability-pair} inequality is sharp. This does not establish sharpness of the Markov-process inference bound, which includes further data-processing and cone relaxations. If supports differ, the divergence is infinite and the inequality follows by a limiting argument. \section{Path reversal, support and the factor of two} \label{app:path} Stationarity gives \begin{equation} \sum_{y\ne x}q^\dagger_{xy} =\frac1{\pi_x}\sum_{y\ne x}\pi_yq_{yx} =-q_{xx}. \end{equation} Thus $Q$ and $Q^\dagger$ have the same escape rates. Their finite-time conditional path laws $\mathsf P_x$ and $\mathsf P_x^\dagger$, both started at $x$, have the same support by reciprocity. The log Radon--Nikodym derivative contains only jump terms, \begin{equation} \log\frac{d\mathsf P_x}{d\mathsf P_x^\dagger} =\sum_{\text{jumps }u\to v}\log\frac{q_{uv}}{q^\dagger_{uv}}. \end{equation} All log ratios on supported edges are finite; finite-state jump counts have finite expectation on bounded intervals. Averaging over $x\sim\pi$ therefore gives \begin{align} \sum_x\pi_x\KL(\mathsf P_x\Vert\mathsf P_x^\dagger) &=\tau\sum_{u\ne v}\pi_uq_{uv} \log\frac{\pi_uq_{uv}}{\pi_vq_{vu}} =\tau\sigma(Q),\\ \sum_x\pi_x\KL(\mathsf P_x^\dagger\Vert\mathsf P_x) &=\tau\sigma(Q^\dagger)=\tau\sigma(Q). \end{align} The two outward half-paths of a stationary two-sided Markov process are conditionally independent given $X_0$; the past half has generator $Q^\dagger$. Applying the same observation/classification channel to the two conditional laws gives $a_x,b_x$ and the final inequality in \eqref{eq:chain}. An event has zero probability under one conditional law exactly when it does under the other, so $H_x>0$. There are finitely many midpoint states, hence $\delta<1$. The total measured interval is $2\tau$, whereas each one-sided conditional divergence contributes $\tau\sigma$. This explains $2/\tau$ in \eqref{eq:main}; substituting the full window length for $\tau$ without adjusting the coefficient would change the claim. \begin{proposition}[Strictness for finite dissipative chains] \label{prop:strictness} For every finite chain and finite positive window in Theorem~\ref{thm:main}, $\sigma(Q)>0$ implies $\sigma(Q)>2\ff(\dc(S))/\tau$. Equality holds exactly at equilibrium. \end{proposition} \begin{proof} The affinity argument in Appendix~\ref{app:affinity} applies to mutually absolutely continuous probability measures as well as finite vectors. For $R=(P+P')/2$ and $U=2\sqrt{pp'}/(p+p')$, it gives $J(P,P')=4\E_R h(U)\ge4h(\E_R U)$, with equality only if $U$ is constant. Each conditional path pair $\mathsf P_x,\mathsf P_x^\dagger$ has a common no-jump atom of equal positive mass $e^{q_{xx}\tau}$; there $U=1$. Hence equality forces identical conditional path laws. If $\sigma>0$, an edge has nonzero stationary affinity, and the event of exactly that jump gives a different likelihood ratio with positive probability for its initial midpoint. At least one conditional affinity inequality is therefore strict. A common classification channel increases affinity by Cauchy--Schwarz. Writing $H_x^{\rm path}$ for the path affinity yields \[ 2\tau\sigma >4\sum_x\pi_x\ff(1-(H_x^{\rm path})^2) \ge4\ff\!\left(\sum_x\pi_x(1-(H_x^{\rm path})^2)\right) \ge4\ff(\dc(S)). \] At equilibrium the identical-half Gram representation gives zero deficit. \end{proof} This is pointwise strictness. It supplies no uniform positive gap over varying generators and does not rule out asymptotic sharpness. \section{Complete positivity: optimization, duality and contraction} \label{app:cone} For completeness, $\CP_m$ is closed. Normalize each nonzero generator $uu^{\mathsf T}$ by its trace. These normalized generators form a compact set, and their convex hull is compact in the finite-dimensional space of symmetric matrices. A convergent sequence of completely positive matrices has bounded traces, and thus has a limit in the same cone. Equivalently, finite-dimensional Carath\'eodory reduction supplies uniformly bounded numbers of normalized factors. For fixed $S\ge0$, the feasible set $\{G\in\CP_m:0\le G\le S\}$ is nonempty and compact. A maximizing $G$ must have $G_{ii}=S_{ii}$, since a positive diagonal deficit can be added as a nonnegative coordinate Gram factor. Its total mass cannot exceed one. If the maximum is one, the nonnegative matrix $S-G$ has zero total mass and hence vanishes. This proves $\dc(S)=0\Longleftrightarrow S\in\CP_m$. Weak duality follows from \begin{equation} -\ip W S=-\ip W G-\ip W{S-G} \le\one^{\mathsf T}(S-G)\one. \end{equation} For the reverse inequality take $0\le t<\dc(S)$ and consider the closed convex set \begin{equation} \mathcal C_t=\{G+E:G\in\CP_m,\ E=E^{\mathsf T}\ge0, \one^{\mathsf T}E\one\le t\}. \end{equation} It is closed because it is the sum of a closed cone and a compact set. Since $S\notin\mathcal C_t$, strict separation gives a symmetric $W$ with $\ip W S<\inf_{C\in\mathcal C_t}\ip W C$. The infimum is finite only if $W$ is copositive. Its value is $t\min(0,\min_{ij}W_{ij})$. Since $S\ge0$, separation is impossible if $W\ge0$ entrywise. Normalize the negative minimum entry to $-1$; then $-\ip W S>t$. Letting $t\uparrow\dc(S)$ proves \eqref{eq:duality}. If $\dc(S)=0$, $W=0$ and weak duality suffice. No strict-positivity assumption on $S$ is used. Finally, if $G$ is a maximizing minorant and $K\ge0$ has row sums one, then $K^{\mathsf T}GK$ is completely positive, is dominated entrywise by $K^{\mathsf T}SK$, and has the same total mass as $G$. It is a feasible minorant for the coarsened matrix and proves \eqref{eq:contraction}. \begin{proposition}[Sharp contraction under matched relabeling] \label{prop:sharp-channel-contraction} Let $S=S^{\mathsf T}\ge0$ have total mass one, and let $K\ge0$ be row-stochastic. Define the row affinities and channel coefficient \[ b_{ij}(K)=\sum_a\sqrt{K_{ia}K_{ja}},\qquad c(K)=1-\min_{i,j}b_{ij}(K)^2. \] Then \begin{equation} \dc(K^{\mathsf T}SK)\le c(K)\dc(S). \label{eq:sharp-channel-contraction} \end{equation} For at least two input categories this is the exact best multiplier over all symmetric nonnegative probability matrices. In particular, pairwise overlap of all channel-row supports implies strict contraction whenever $\dc(S)>0$. \end{proposition} \begin{proof} For probability vectors $p,q$, let $A=(pq^{\mathsf T}+qp^{\mathsf T})/2$ and $v_a=\sqrt{p_aq_a}$. AM--GM gives $vv^{\mathsf T}\le A$ entrywise. Every completely positive minorant $H\le A$ is positive semidefinite, and hence \[ \one^{\mathsf T}H\one \le\left(\sum_a\sqrt{H_{aa}}\right)^2 \le\left(\sum_a\sqrt{p_aq_a}\right)^2. \] The rank-one minorant attains this bound. Now choose an optimal minorant $G\le S$, and put $E=S-G$. Diagonal saturation gives $E_{ii}=0$. For rows $i,j$ of $K$, define $v^{ij}_a=\sqrt{K_{ia}K_{ja}}$. The completely positive matrix \[ B=\sum_{i