\section{A count-only obstruction beyond the positive-determinant regime}\label{sec:negative} This section allows any finite number of hidden states and does not assume $\det R>0$ or prescribe $L$. The direct and excursion rates appearing in logarithms are assumed positive. The affinity argument is an application of standard relative-entropy contraction; its matrix representation exposes the relevant determinant obstruction. \begin{theorem}[Occurrence-determinant bound]\label{thm:negative} For every finite irreducible reciprocal chain under the stated observation convention, \begin{equation} \sigma(Q)\ge B(d)+(R_{12}+R_{21}) \left[\log\frac{R_{12}R_{21}}{R_{11}R_{22}}\right]_+, \qquad [x]_+=\max(x,0). \label{eq:negative-bound} \end{equation} In particular, if $\det R<0$, every compatible model, irrespective of its finite hidden dimension, has dissipation strictly greater than $B(d)$ by the displayed amount. \end{theorem} \begin{proof} For a hidden state $h$, let $d_h=\sum_{y\ne h}n_{hy}$ be its total stationary departure flux. It is positive. Define the symmetric hidden matrix and visible-to-hidden matrix \begin{equation} J_{hh}=d_h,\quad J_{hk}=-\sqrt{n_{hk}n_{kh}}\ (h\ne k), \qquad Z_{ih}=\sqrt{n_{ih}n_{hi}}.\label{eq:geometric-killed} \end{equation} These matrices use flows, not transition rates. By stationarity and the arithmetic--geometric mean inequality, \begin{equation} d_h-\sum_{k\in H\setminus\{h\}}\sqrt{n_{hk}n_{kh}} \ge\frac12\sum_{i\in V}(n_{hi}+n_{ih})\ge0. \label{eq:geometric-dd} \end{equation} Every hidden connected component has an edge to a visible state, because the full reciprocal graph is connected. Hence on each such component at least one row is strictly diagonally dominant. Explicitly, for a vector $u$, \[u^TJu=\sum_{\{h,k\}\subset H}\sqrt{n_{hk}n_{kh}}(u_h-u_k)^2 +\sum_h\left(d_h-\sum_{k\ne h}\sqrt{n_{hk}n_{kh}}\right)u_h^2. \] Vanishing forces $u$ constant on each hidden component and zero at a strictly dominant row. Thus $J$ is positive definite. Let $\gamma=(i,h_1,\ldots,h_m,j)$ be an excursion jump sequence, allowing repeated hidden states but no consecutive repeats. After the holding times are integrated out, its stationary rate is \begin{equation} w(\gamma)=\frac{n_{ih_1}n_{h_1h_2}\cdots n_{h_{m-1}h_m}n_{h_mj}} {d_{h_1}\cdots d_{h_m}}.\label{eq:path-rate-discrete} \end{equation} The holding-time factors in the forward/reverse ratio cancel before integration, so keeping this jump sequence keeps its path-entropy contribution. Reversal, denoted $\bar\gamma$, is a bijection between opposite endpoint classes. Summing their geometric means yields \begin{equation} G_{ij}=\sum_{\gamma:i\to H\to j}\sqrt{w(\gamma)w(\bar\gamma)} =(ZJ^{-1}Z^T)_{ij}.\label{eq:affinity-Gram} \end{equation} To justify the sum, write $J=D-A$ with $D=\diag(d_h)$ and $A\ge0$ symmetric. The normalized matrix $D^{-1/2}AD^{-1/2}$ has spectral radius below one. Indeed positive definiteness bounds its largest eigenvalue below one, and nonnegativity makes that eigenvalue its spectral radius. The Neumann series for $J^{-1}$ then expands exactly into the successive hidden paths in \eqref{eq:path-rate-discrete}. The matrix $G$ is positive semidefinite. Cauchy--Schwarz in each return-path class gives $G_{ii}\le R_{ii}$, since reversal preserves the total return rate. Consequently \begin{equation}G_{12}\le\sqrt{G_{11}G_{22}}\le\sqrt{R_{11}R_{22}}.\label{eq:affinity-upper}\end{equation} On the set of $1\to H\to2$ sequences, normalize forward and reversed rates as $p=w/R_{12}$ and $q=w\circ\mathrm{rev}/R_{21}$. Their affinity is \[A(p,q)=\sum_\gamma\sqrt{p(\gamma)q(\gamma)} =G_{12}/\sqrt{R_{12}R_{21}}. \] Jensen's inequality under $p$ gives $D(p\Vert q)\ge-2\log A(p,q)$; the same inequality holds with $p,q$ reversed. These sums are finite in entropy for every fixed admissible model, by the stationary edge-counting argument used for \eqref{eq:H5}. Applying that argument to jump sequences gives \[\sigma(Q)-B(d)\ge R_{12}D(p\Vert q)+R_{21}D(q\Vert p) \ge-2(R_{12}+R_{21})\log A(p,q). \] Use \eqref{eq:affinity-upper} when its ratio is below one, and the original nonnegative entropy excess otherwise. This proves \eqref{eq:negative-bound}. \end{proof} \subsection{An exact negative-determinant example} Take two independent exponential channels \begin{equation} M_1=\begin{pmatrix}1&4\\1/4&1\end{pmatrix},\qquad M_2=\begin{pmatrix}1&1/4\\4&1\end{pmatrix},\qquad t_1=t_2=1/50. \end{equation} Both matrices are positive rank one. The channel construction gives \begin{equation} R=\begin{pmatrix}2&17/4\\17/4&2\end{pmatrix},\quad L=R/50, \quad P=(3/8,3/8,1/4),\quad F_{12}=F_{21}=1. \end{equation} The two hidden masses are $1/8,1/8$, all required rates are positive, and $B(d)=0$. Nevertheless, \begin{equation} I_\infty(d)\ge\frac{17}{2}\log\frac{289}{64}>0. \end{equation} The displayed model has $\sigma=(15/2)\log16$, which is an upper witness for the present inequality. Section~\ref{sec:all-regimes} strengthens the inequality and proves that this witness is optimal. For this model, $J=(25/4)I$ and all entries of $Z$ equal $5/2$, so $G$ is the matrix with all entries equal to two. Thus the Gram constraint is exact here, even though the relative-entropy inequality need not be saturated. Equal cross means by themselves do not imply equality of the unrestricted infimum with $B(d)$ outside the positive-determinant regime. The present coarse affinity inequality alone does not decide reduced safety; the equality classification in Section~\ref{sec:all-attainment} does so.