\section{The zero-gap finite far envelope}\label{app:middlefar} Continue with the coefficients \eqref{mid:valuemultipliers}. Put $m_0=3/4$, $u_0=7/8$, $\delta_0=1/8$, and define \begin{equation}\label{mf:F} \mathcal F_0(p)=N+\lambda_*\bar Q+\mu_*\bar H +T_*/8+(1/10)m_0^{-p}[C_*]_+. \end{equation} Its contributions from $S,q,H$ are explicitly \begin{align*} N={}&u_0(1+pa-Z)+m_0G_u(u_0)+I_G(x) +(m_0/2)G_u(x)-(m_0^2/12)G_u'(x),\\ &a=189/800,\quad z=511/800,\quad x=13/8,\quad y=x+1=21/8,\\ I_G(x)={}&zx^{-p}+ay^{-p}-\int_x^y t^{-p}\dd t,\\ \bar Q={}&s[p+1-Z+13M_2/72-M_3/72+1063/1296000]\\ &+Z-1-p/8+(1/8-3p/4)r+M_2/128+1/2048,\\ &s=(4/3)^{p-1},\quad r=(8/7)^p,\\ \bar H={}&-(9/128)[I(13/16)^{-p-2}+I_x(15/16)^{-p-2}]. \end{align*} The integral form of $I_G$ also applies at $p=1$, where its last term is $-\log(y/x)$. \begin{lemma}[Complete far value]\label{mid:farvalue} At $u=u_0$ and $p\in[1,9/5]$, the function \eqref{mf:F} bounds \[ \Phi(m_0,0)+\mathcal C(p,\delta_0) +\tfrac1{10}m_0^{-p}[C_*]_+\leq\mathcal F_0(p). \] It satisfies $\mathcal F_0(p)<-.00052-.009(P-p)$ for every $1\leq p\leq9/5$. \end{lemma} \begin{proof} We first justify the three envelopes rather than treating them as formal expressions. For $N$ the signed Euler upper is valid because \[ E_2(z)=1/z+1/2+z/12-1/(1-e^{-z}) \] is positive increasing. Its derivative sign reduces to $(z^2-12)(\cosh z-1)+6z^2>0$, whose coefficients through degree four vanish and whose later even coefficients are positive. The Laplace density of $G_u(x)$ has one crossing, from negative to positive, by Lemma~\ref{mid:signedpacking}. Its integral is positive: trapezoidal convexity on $[x,y]$ gives \[ G_u(x)\ge p x^{-p-1}\big[z-1/2+(a-1/2)(x/y)^{p+1}\big]>0, \] since $x/y=13/21$ and $p\ge1$. The difference between the three Euler terms in $N$ and the exact following sum is the integral of this density times $m_0E_2(m_0s)$. The positive increasing multiplier preserves positivity across the crossing. For $\bar Q$, use $F(U)=\sum_{j\geq1}A_U(j)$. Its first two jets at one are \[ F(1)=p+1-Z,\quad F'(1)=M_2,\quad F''(1)=M_2-M_3. \] On $1\leq U\leq7/6$, $F'''(U)<1.063$. Indeed the full differentiated series is bounded above by \[ (p)_3\sum_{n\geq2}n^{-p-3}[7(p+3)/(6n)-4/3]. \] All $n\geq5$ terms are negative; the other three positive parts increase with $p$ and at $P$ are less than $1.008347$, $.0524$, $.001646$. For the adjacent left stream, the entire third derivative is bounded by \[ (p)_3(15/8)^{-p-3}+p(p+1)(15/8)^{-p-2}<3/2. \] Taylor's integral formula at $\delta_0=1/8$, $U-1=1/6$ now gives exactly $\bar Q$. Finally the diagonal identity for $H$ in Lemma~\ref{mid:jointstrip}, the elasticity bound for $b_0$, and Jensen's inequality on the intervals $[m_0,u_0]$, $[u_0,1]$ give the negative upper $\bar H$. Their interval midpoints are $13/16$ and $15/16$, and $m_0^2\delta_0=9/128$. The complete exponent curvature bounds are \begin{equation}\label{mf:curvatures} N''>.049,\qquad \bar Q''>-.25,\qquad \bar H''>-.51. \end{equation} For $N$, separate the five printed terms. The first two together have second derivative greater than $.75(.358)-.875(.125)=.159125$. Here $G_{pp}(u_0)>.358$: the fourth derivative is positive and at $P$, $G_{pp}(u_0)>.3598$, $G_{ppp}(u_0)<.002$. To check these claims, for an affine power use \[ \partial_p^k[(Ap+B)v^{-p-r}] =v^{-p-r}\big[(-\log v)^k(Ap+B)+kA(-\log v)^{k-1}\big]. \] At $u_0<1$ this gives a positive fourth derivative for $(pz/u_0-1)u_0^{-p}$ because $4z/u_0>|\log u_0|$. At $v=u_0+1$ the other block $(1+pa/v)v^{-p}$ also has positive fourth derivative because $\log v(1+a/v)>4a/v$. The endpoint bounds then give $G_{pp}(u_0)>.3598-.8(.002)>.358$. For $G_{pp}(x)$, put $\ell=\log x$, $L_y=\log y$, $A_x=z\ell^2/x$ and $A_y=a L_y^2/y$. Its derivative is \[ x^{-p}[\ell^3+(3-p\ell)A_x] -y^{-p}[L_y^3+(p L_y-3)A_y]. \] On each of the four intervals $[l,h]=[1+j/5,1+(j+1)/5]$, the second bracket increases, the first decreases, and the second is positive. The difference after division by $x^{-p}$ is at most $\ell^3+(3-l\ell)A_x-(x/y)^h[L_y^3+(lL_y-3)A_y]<-.007$. Thus $G_{pp}(x)\ge G_{pp}(x)|_{p=P}>-.029$. For $I_G^{(3)}>0$, the log-concave function $v^{-p}\log^3v$ on $[x,y]$ has endpoint ratio between three and five. The integral is at least the logarithmic mean of the two endpoint values, which exceeds $.4$ times their sum on this ratio interval. This pays the two endpoint terms, since $.4-z+3(.4-a)>0$. At $p=1$ the second derivative exceeds \[ .393(.485)^2+.09(.965)^2 -[(.966)^3-(.485)^3]/3>-.087. \] For the four terms of $(-G_x)_{pp}$, write $D(v)=v^{-p-2}[(\log v)^2p(p+1)-2\log v(2p+1)+2]$. The terms and their lower bounds are \[ \begin{array}{c|r} zD(x)&-.191614\\ \partial_p^2(-p x^{-p-1})&.1399\\ \partial_p^2(p y^{-p-1})&-.146113\\ aD(y)&-.03318 . \end{array} \] The two affine terms attain their minima at $P$ and $1$, respectively, by the third-derivative formula. For the two quadratic terms, evaluate the displayed products by interval substitution on intervals of length $1/20$ starting at one. The same power rule checks the four intervals for $G_{ppp}(x)$ and the endpoint derivatives above. For reproduction, enclose $\log v$ by the first 24 terms of the odd logarithm series of Appendix~\ref{app:escort} and its geometric tail, using inversion for $v<1$. For $v^{-p-r}$, use the exponential series through degree 32 at the appropriate endpoints of the exponent interval. For a negative exponent enclose the reciprocal of the positive exponential. At $t\ge0$ the omitted tail is at most $t^{33}/[33!(1-t/34)]$. Rounding interval operations outwards to multiples of $10^{-18}$ gives the stated strict bounds. Thus $(-G_x)_{pp}>-.25$, and $N''>.159125-.087-.375(.029)-.046875(.25)>.049$. For $\bar Q$, keep its common $Z$ before differentiating. Lemma~\ref{mid:logmoments} gives the following lower bounds for its second-derivative terms, in order: \[ \begin{array}{c|r} (1-s)Z''&-.03235\\ 2\log(4/3)s(1-Z')&.1355117568\\ s\log^2(4/3)(p+1-Z)&.029240995\\ {[(1/8-3p/4)r]''}&-.2836510392\\ (13s/72+1/128)M_2''&-.0940383333\\ (13s/36)\log(4/3)M_2'&.0020727778\\ (13s/72)\log^2(4/3)M_2&.0059488722\\ -(sM_3)''/72&-.01130301696 \end{array} \] The omitted constant-times-$s$ derivative is positive. Their sum exceeds $-.248568$, proving the stated bound. For $\bar H$, differentiation of the two products uses $P_m''<1.716$, $P_m'<2.892$, $I<2.8$, $I_x''<.357$, $I_x'<.247$, $I_x<.5$, the powers $<2.21,1.28$ and logarithms $<.208,.065$. These give $\bar H''>-.51$. The bound $I_x<.5$ follows directly from $e^sC(s)\leq1/2+s$ and $b_0(s)\leq s^2/2$: $I_x\leq.25(p)_2 2^{-p-2}+.5(p)_3 2^{-p-3}<.435456$. The remaining positive part in \eqref{mf:F} has second derivative at least $-.1(1.679)[.095+2(.288)(.069)]>-.023$. This is a distributional lower curvature: the first derivative jumps upward at an activation of $[C_*]_+$. Since $T_*$ is convex, \[ \mathcal F_0''>.049-.178(.25)-.017765(.51)-.023>-.03. \] At $P$, the printed finite expressions and the full moment enclosures give $N<-.042376$, $\bar Q<.26363$, $\bar H<-.4438$. Since $C_*(P)=0$ and $T_*(P)/8<.0028$, \[ \mathcal F_0(P)<-.042376+.178(.26363)-.017762(.4438)+.0028 <-.00052. \] At one, direct powers and logarithms give $N<-.115$, $\bar Q<.232$, $\bar H<-.13$, $C_*(1)<.008606$, $T_*(1)/8=23/1176$. Consequently $\mathcal F_0(1)<-.054$. For $\theta=(p-1)/(4/5)$ the semiconvex chord is \[ \mathcal F_0(p)<-.054(1-\theta)-.00052\theta +.0096\theta(1-\theta) \leq-.00052-.009(P-p). \] This proves the full finite value, including every activation contact. \end{proof}