\section{A fixed reduction bias can be hidden at finite observation times} \label{sec:first-moment-non-testability} The singular channels approaching the first-moment infimum also obstruct uniform finite-time detection. Rare mean-changing contamination is classical \cite[Theorem~1 and its proof]{BahadurSavage1956}; here it is realized by stationary reciprocal chains at fixed prescribed data. A positive reduction bias and determinant, even with a fixed hidden-state budget, do not ensure uniform detection power from continuously observed stationary labels in a fixed time unit. Write \[ \mathfrak b(d)=I_2(d)-I_\infty(d) \] only when the two-hidden-state fitting class is nonempty. \begin{proposition}[Fixed-bias finite-window non-testability] \label{prop:fixed-bias-non-testability} There are two fixed strictly positive consistent data $d^{\mathrm S},d^{\mathrm U}$, both two-hidden-state feasible and sharing the same $P,F,R$, such that \[ \det R=\frac{21}{4},\qquad \mathfrak b(d^{\mathrm S})=0,\qquad \mathfrak b(d^{\mathrm U})\ge g_*:=\frac{5(17-12\sqrt2)}{18}>0. \] For each $0<\varepsilon\le1$ there are reciprocal irreducible chains $Q^{\mathrm S}_\varepsilon,Q^{\mathrm U}_\varepsilon$, each with four hidden states, realizing their respective data exactly and having a common finite upper bound on transition rates. If $\mathsf P^\ell_{\varepsilon,T}$ is the law of their stationary observed label trajectory on $[0,T]$, then for finite constants $c_0,c_1$ independent of $\varepsilon,T$, \begin{equation}\label{eq:fixed-bias-TV} \left\|\mathsf P^{\mathrm S}_{\varepsilon,T} -\mathsf P^{\mathrm U}_{\varepsilon,T}\right\|_{\rm TV} \le c_0\varepsilon+c_1\varepsilon T. \end{equation} In particular, for every fixed $T<\infty$ this distance tends to zero. Fix $0<\alpha<1$ and a possibly randomized test $0\le\phi_T\le1$ declaring an unsafe reduction. If $\E_Q\phi_T\le\alpha$ for every chain with safe two-hidden-state-feasible data, then \begin{equation}\label{eq:fixed-bias-power} \inf_{Q:\,d(Q)=d^{\mathrm U}}\E_Q\phi_T\le\alpha. \end{equation} The assertion remains valid when the common exact $P,F,R$ are supplied to the test. \end{proposition} \begin{proof} Use the unsafe datum of Appendix~\ref{feasible-but-unsafe-reduced-data}, and denote its moment matrix by $L^{\mathrm U}$. Thus \[ R=\begin{pmatrix}9/2&3\\2&5/2\end{pmatrix},\qquad L^{\mathrm U}=\begin{pmatrix}1/20&1/25\\3/100&9/200\end{pmatrix}, \] \[ P=(2/5,87/200,33/200),\qquad F_{12}=1,\quad F_{21}=2. \] Its two-channel witness and Theorem~\ref{thm:H3} give the stated lower bound $g_*$. Since $\one^TR\one=12$, set \[ t_0=\frac{P_H}{\one^TR\one}=\frac{11}{800},\qquad L^{\mathrm S}=t_0R. \] The safe datum has a two-channel witness \[ N_1=\begin{pmatrix}9/2&3\\2&4/3\end{pmatrix},\qquad N_2=\begin{pmatrix}0&0\\0&7/6\end{pmatrix}, \] both with residence time $t_0$. Their hidden entropy is $\log(3/2)$; the direct edge contributes $\log2$. Hence the witness attains $B(d^{\mathrm S})=\log3$, and Theorem~\ref{thm:H1} proves safety. For both data use the exact channel construction of Section~\ref{exact-first-moments-with-arbitrarily-little-extra-dissipation}, with $\delta=1/1000$: \begin{equation}\label{eq:fixed-bias-channels} R_f^\ell=\frac{R-\varepsilon L^\ell}{1-\delta\varepsilon}, \qquad S^\ell=\frac{L^\ell-\delta R}{1-\delta\varepsilon}, \qquad \ell\in\{\mathrm S,\mathrm U\}. \end{equation} These matrices are positive for $0<\varepsilon\le1$, and $\det R_f^\ell>0$. In fact, \[ \det(R-\varepsilon L^{\mathrm U}) =\frac{21}{4}(1-\varepsilon/100)(1-\varepsilon/50),\qquad \det(R-\varepsilon L^{\mathrm S}) =\frac{21}{4}(1-11\varepsilon/800)^2. \] Split $R_f^\ell$ into the two fast matrices in \eqref{eq:H9}, both with residence time $\delta$. For the slow factorization take \[ Y=\begin{pmatrix}99/100&1/100\\1/100&99/100\end{pmatrix}, \qquad X^\ell=S^\ell Y^{-1},\qquad S_j^\ell=X^\ell_{\cdot j}Y_{j\cdot}\quad(j=1,2). \] Every entry of $X^\ell$ is positive: each row's smaller entry in $L^\ell-\delta R$ exceeds $1/100$ of that row's sum. Give occurrence matrix $\varepsilon S_j^\ell$ residence time $1/\varepsilon$. The identities \[ R_f^\ell+\varepsilon S^\ell=R,\qquad \delta R_f^\ell+S^\ell=L^\ell \] prove exact agreement with $R,L$. The channel construction and consistency supply the prescribed masses, flows and stationarity. All supported edges are reciprocal, and the direct visible edge makes the chain irreducible. Fast exit rates are $1000$, slow exit rates are $\varepsilon$, and visible departure rates are fixed by $P,F,R$. Thus all transition rates have a common finite upper bound. It remains to compare the stationary \emph{observed} paths. The aggregate stationary slow mass is the same in both chains: \begin{equation}\label{eq:fixed-bias-slow-mass} m_{\rm slow}(\varepsilon)=\one^TS^\ell\one =\frac{P_H-\delta\one^TR\one}{1-\delta\varepsilon} =\frac{153/1000}{1-\varepsilon/1000}. \end{equation} The two fast-state masses and rates are smooth in $\varepsilon$, with common positive limits at zero; their differences are $O(\varepsilon)$. Slow entrance rates are $O(\varepsilon)$. Couple the stationary initial categories \[ 1,\ 2,\ f_1,\ f_2,\ \{\hbox{slow states}\}. \] Their probabilities differ by $O(\varepsilon)$, because visible and aggregate slow masses agree exactly. If the common initial category is slow, both observed labels stay hidden until a slow exit; since each residual slow exit time is exponential of rate $\varepsilon$, the two exit times can be coupled to coincide, and the probability of this exit by $T$ is at most $\varepsilon T$. The slow microscopic states need not agree. If the initial common category is visible or fast, couple each transition using the minimum of the two rates, declaring failure at an unmatched transition or any slow entrance. Before failure the states in this common subsystem agree, and the failure intensity is at most a constant times $\varepsilon$. Uniform constants follow from the positive fast limits and bounded rational coefficients on $[0,1]$. Initial mismatch and these failure probabilities prove \eqref{eq:fixed-bias-TV} by the coupling inequality. Finally the difference between the expectations of a $[0,1]$-valued test is at most total variation. Apply the uniform safe level bound and let $\varepsilon\downarrow0$ to obtain \eqref{eq:fixed-bias-power}. \end{proof} The reducible endpoint $\varepsilon=0$ serves only to bound coefficients and compare limits. Pointwise eventual detection for a fixed chain remains possible; arbitrary growing horizons are not covered. Procedures stopping by $T$ are covered, since their output is measurable from the path on $[0,T]$. The fixed bias compares fitting infima, not a fitted value with the actual generator's entropy.