\documentclass[11pt]{amsart} \IfFileExists{lmodern.sty}% {\usepackage[T1]{fontenc}\usepackage{lmodern}\usepackage{microtype}}% {\usepackage[expansion=false]{microtype}} \usepackage{amssymb,mathtools} \usepackage{booktabs,array} \usepackage{enumitem} \emergencystretch=2.5em \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref} \newcommand{\FormalizationURL}{https://github.com/slitsey/catalan-irrationality-measure} \newcommand{\FormalizationLink}{\url{\FormalizationURL}} % Numerical constants: decimal truncations (or outward roundings) of rational numbers of the formal proof. \newcommand{\Mustar}{269.78696} \newcommand{\Rho}{278.376313578} \newcommand{\Ctwo}{3.596893} \newcommand{\Sigm}{0.33871} \newcommand{\Cost}{3.935603} \newcommand{\Psiw}{50.111339578} \newcommand{\Czero}{33.052} \newcommand{\Dcirc}{3.2506} \newcommand{\Utarget}{2.255858836} \newcommand{\Gain}{24.64} \newcommand{\Uzero}{2.2631834} \newcommand{\NormCap}{0.77363} \newcommand{\SupX}{-0.96450499} \newcommand{\SupY}{-1.57499400} \newcommand{\FieldCap}{-2.53949899} \newcommand{\ResCap}{-0.0631062} \newcommand{\EtaTot}{54.68745} \newcommand{\LamCrit}{0.9699101} \newcommand{\Kcoef}{160} \newcommand{\Ngeo}{75} \newcommand{\Ncomplex}{41} \newcommand{\Mmargin}{2.986\cdot10^{-6}} \newcommand{\Decay}{6447.4755} \newcommand{\DecayOA}{6422.8355} \newcommand{\EtaExact}{54.68853} \newcommand{\PayL}{53.0000596} \newcommand{\PayR}{53.0419104} \newcommand{\AlphaOne}{51/500} \newcommand{\Chi}{22.06076} \newcommand{\ChiLean}{22.0607} \newcommand{\LamDen}{6392.7870} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \setcounter{tocdepth}{1} \newcommand{\Z}{\mathbb Z} \newcommand{\Q}{\mathbb Q} \newcommand{\R}{\mathbb R} \newcommand{\N}{\mathbb N} \newcommand{\C}{\mathbb C} \newcommand{\F}{\mathbb F} \newcommand{\rank}{\operatorname{rank}} \newcommand{\cV}{\mathcal V} \newcommand{\cR}{\mathcal R} \newcommand{\cA}{\mathcal A} \newcommand{\cL}{\mathcal L} \newcommand{\cK}{\mathcal K} \newcommand{\cD}{\mathcal D} \newcommand{\rev}{\operatorname{rev}} \newcommand{\lean}[1]{\texttt{#1}} \begin{document} \title{An irrationality measure for Catalan's constant} \author{Sean Litsey} \email{sean.litsey.math@gmail.com} \date{November 8, 2026} \subjclass[2020]{Primary 11J82; Secondary 11J72, 15A15, 31A15, 68V20} \keywords{Catalan's constant, irrationality measure, determinants, corank modulo $p$, Andr\'eief's identity, logarithmic energy, Pick matrices, Nevanlinna--Pick interpolation, Gauss quadrature, Lean} \begin{abstract} Let $G=\sum_{j\ge0}(-1)^j(2j+1)^{-2}$ be Catalan's constant. We prove that $|G-a/b|>b^{-270}$ for all integers $a$ and all sufficiently large integers $b$. Hence the irrationality exponent of $G$ is at most $270$. The proof refines OpenAI's recent proof that $G$ is irrational. That proof compares two bounds for a large determinant whose entries are linear forms in $1$ and $G$. The lower bound comes from the denominators of the entries, and the upper bound from an integral representation. We improve both. On the arithmetic side we show that the matrix of rational parts has large corank modulo $p$ for the primes in explicit intervals, which lowers the denominators. On the analytic side we bound a determinant that the energy method treats trivially. A phase inequality for Pick matrices, proved by Nevanlinna--Pick interpolation, compares it with a determinant of a simpler kind. That determinant is bounded by a finite Gram inequality and two logarithmic gases. Gauss quadrature at Chebyshev nodes of many short intervals extends the upper bound to an exponentially small neighbourhood of $G$, as a measure requires. The proof is formalized in the Lean proof assistant and kernel-checked. \end{abstract} \maketitle \tableofcontents \section{Introduction}\label{sec:intro} \subsection{The result} Catalan's constant \[ G=\sum_{j=0}^\infty\frac{(-1)^j}{(2j+1)^2}=L(2,\chi_{-4})=0.91596\,55941\,77219\ldots \] is the value at $s=2$ of the Dirichlet $L$-function of the nontrivial character modulo $4$. Values of $L$-functions at integers where no formula in terms of $\pi$ is available are notoriously hard to study. Ap\'ery proved $\zeta(3)$ irrational in 1978 \cite{Apery79}, and Calegari, Dimitrov and Tang proved the Calegari--Dimitrov--Tang constant $L_{-3}=L(2,\chi_{-3})$ irrational in 2024 \cite{CDT2024}. For $G$ the classical methods produce linear forms that do not decay fast enough, and its irrationality remained open. In October 2026 OpenAI \cite{OpenAI2026} proved that $G$ is irrational, with a new argument based on a large determinant, and verified the proof in Lean. The present paper turns that argument into an irrationality measure. The \emph{irrationality exponent} $\mu(\xi)$ of a real number $\xi$ is the infimum of the real numbers $\mu$ such that $|\xi-a/b|>b^{-\mu}$ for all integers $a$ and all sufficiently large integers $b$. Every irrational number has $\mu(\xi)\ge2$, and almost every real number has $\mu(\xi)=2$. An \emph{irrationality measure} is an explicit upper bound for $\mu(\xi)$. The exponent does not change under $\xi\mapsto r\xi+r'$ with $r,r'\in\Q$ and $r\ne0$. \begin{theorem}\label{thm:main} For all integers $a$ and all sufficiently large integers $b$, \[ \left|G-\frac ab\right|>b^{-270}. \] In particular $\mu(G)\le270$. \end{theorem} The proof uses the prime number theorem. The theorem and every result it depends on have been formalized in the Lean~4 proof assistant on top of OpenAI's formalization, and the proof is kernel-checked. The formal proof has no unproved steps and uses only the standard axioms (Section~\ref{sec:formal}). The use of AI in this work is disclosed at the end of the paper. \subsection{Background} \emph{Catalan's constant.} Irrationality proofs for values of $L$-functions at integers usually follow Ap\'ery's proof for $\zeta(3)$. One constructs linear forms $q_n\xi-p_n$ with integer coefficients that tend to zero faster than the coefficients grow. If $|q_n\xi-p_n|=e^{-\alpha n+o(n)}$ and $|q_n|=e^{\beta n+o(n)}$ with $\alpha>0$, then $\xi$ is irrational and $\mu(\xi)\le1+\beta/\alpha$. Beukers \cite{Beukers79} obtained Ap\'ery's forms for $\zeta(2)$ and $\zeta(3)$ as integrals of Legendre-type polynomials. Hata \cite{Hata93} showed how much is gained by dividing the forms by the primes that divide all their coefficients, an idea that goes back to Chudnovsky. Rhin and Viola \cite{RhinViola96,RhinViola01} found finite groups acting on families of such integrals. The invariance yields a large common factor of the coefficients, and this \emph{arithmetic method} gave the best measures for $\zeta(2)$ and $\zeta(3)$ for many years. Marcovecchio \cite{Marcovecchio09} applied it to $\log2$, and Zeilberger and Zudilin \cite{ZeilbergerZudilin20} to $\pi$, using integrals of Salikhov \cite{Salikhov08}. For Catalan's constant the analogous linear forms exist but do not decay fast enough. Rivoal and Zudilin \cite{RivoalZudilin2003} proved that one of the numbers $\beta(2),\beta(4),\ldots,\beta(2m)$ is irrational for suitable $m$, where $\beta$ is the Dirichlet beta function. This does not single out $G=\beta(2)$. Fischler \cite{Fischler20} and Kumar and Tosi \cite{KumarTosi26} obtained lower bounds for the number of irrational values among $\beta(2),\beta(4),\ldots,\beta(2m)$. These bounds again do not single out $G$. Zudilin \cite{Zudilin2003} found an Ap\'ery-like recurrence for $G$, whose solutions give linear forms that tend to zero too slowly. Related constructions and their limits are described in \cite{Zudilin02Cheb,Rivoal06,KrattRivoal08,KrattZudilin19,Zudilin19Survey}. Nesterenko \cite{Nesterenko16} developed a general construction of such approximations and compared their quality. Rational approximations to $G$ with $|G-p_n/q_n|\le q_n^{-0.52}$ follow from these works. Eskandari, Murty and Nemoto \cite{EMN2025} construct a mixed motive with period $G$ and obtain linear forms in $1$ and $G$ from it. Eskandari \cite{Eskandari2026} raises the exponent $0.52$ to $0.62$. An exponent below $1$ does not imply irrationality. \emph{Recent results.} Since September 2026 a series of new results in this area has appeared, several of them obtained with AI systems. OpenAI proved $\mu(\pi)=2$, with a Lean formalization \cite{OpenAIPi2026}. Liu, Jiang and Zhang proved $\mu(\log r)=2$ for every rational $r>0$ with $r\ne1$ \cite{LiuJiangZhang26}. Chudnovsky adapted OpenAI's method to periods of elliptic curves and proved $\mu(\Omega)=2$ for the lemniscate period $\Omega$ \cite{Chudnovsky26}. Giraudin lowered the bound for $\mu(\pi^2)$ to $5.0193784$ with two hypergeometric constructions of Zudilin \cite{Zudilin14} at new parameters \cite{Giraudin26}, and the bound for $\mu(\zeta(3))$ to $5.5138800$ \cite{Giraudin26z3}. Leibtag, Barkan, Kalisch, Weinbaum, Shalyt and Kaminer \cite{LBKWSK26} showed that the classical constructions are one-dimensional trajectories in conservative matrix fields. Searching over other trajectories, they obtained $\mu(\pi^2)<4.62675$ and $\mu(\zeta(3))<5.4197$, and further Ap\'ery-like approximations to $G$, the best of which still has a negative approximation exponent. Prellberg \cite{Prellberg26} proved that $\mathrm{Li}_2(-1/4)$, $\mathrm{Li}_2(1/5)$ and $\mathrm{Li}_2(-1/3)$ are irrational. Fauzan proved that $\zeta(5)$ is irrational, using Hankel determinants \cite{Fauzan2026}. Fasten \cite{Fasten26}, Firsching \cite{Firsching26} and Romik \cite{Romik26} verified the irrationality statement in Lean. Romik's formalization takes the prime number theorem as an axiom. Fauzan's argument also gives an irrationality measure. Using its results, Giraudin proved $\mu(\zeta(5))\le171$ \cite{Giraudin26z5}. Lai combined this method with simultaneous approximations to show that $1$, $\zeta(3)$ and $\zeta(4)$ are linearly independent over $\Q$ \cite{Lai2026}. For $L_{-3}$ the effective holonomy bounds of Calegari, Dimitrov and Tang give $\mu(L_{-3})\le24781$ \cite[Thm.~5.1]{CDT2025}. \emph{Determinant methods.} OpenAI's argument \cite{OpenAI2026} has a different shape. Instead of a sequence of linear forms it uses one determinant $\Delta_N$ of size $n=48N$, whose entries are rational linear combinations of $1$ and $G$. The entries are double integrals against the Cauchy kernel $1/(1-ts)$. If $G$ were rational, the denominators of the entries would force a lower bound $|\Delta_N|\ge e^{-(\Lambda+o(1))N^2}$ along $N=p$ prime. On the other hand, Andr\'eief's identity \cite{Forrester2018Andreief} writes $\Delta_N$ as an integral of a product of determinants, and an estimate for a discrete logarithmic energy gives $|\Delta_N|\le e^{-(D+o(1))N^2}$. OpenAI proves $D>\Lambda$, with a margin of about $10^{-4}$ per unit of $n^2$, and proves that $\Delta_p\ne0$ by reduction modulo $p$. Determinants of linear forms had been used for irrationality before, in Zudilin's criterion with Hankel determinants \cite{Zudilin}. The first irrationality measure for $G$ was obtained by Jha \cite{Jha2026}, who proved $\mu(G)\le10^8$. Carreira \cite{Carreira2026} treated the five integer parameters of the construction as free, stated conditions under which OpenAI's argument goes through, and proved $\mu(G)<2294$. He also expressed the best constant of the energy estimate as the value of an equilibrium problem for two measures, and in \cite{Carreira2026b} he records attempts to carry the method to $\beta(4)$. Krisman proved that one of $\beta(4),\beta(6),\ldots,\beta(36)$ is irrational, with a Lean formalization \cite{Krisman26}. Bojnal \cite{Bojnal2026} and Xu \cite{Xu2026} obtained $\mu(G)<2402$ and $\mu(G)\le727.03$ at other parameters. Xu's certificate is numerical and not formalized. Davis \cite{Davis2026} reached $\mu(G)\le528.23288$ with new quadrature nodes, the zeros of $T_K(P(t))$ for a fixed polynomial $P$. The MultiAgency agent swarm regenerated $P$ and reached $\mu(G)\le528.2315$ \cite{MultiAgency2026}. Both results are formalized in Lean. Our nodes are different, but we use two formal lemmas of \cite{Davis2026} on the weights of nodal polynomials in Section~\ref{sec:real}. The other constructions are independent of the ideas of the present paper. In the third version of \cite{Carreira2026}, Carreira bounds the generalized Vandermonde determinant by one term of the dual Cauchy identity, using the positivity of Schur polynomials. His bound carries the factor $\prod_l(1+\tau s_l)^m$ with $\tau=0.73$, of the same form as in Lemma~\ref{lem:binomvand}, and he obtains $\mu(G)\le280.59$. His certificate is computer-assisted and not formalized. A preprint of Sun \cite{Sun2026} claimed the irrationality of $G$, but Wachs \cite{Wachs2026} showed that its main estimate is not justified. \subsection{The companion paper} The companion paper \cite{Litsey2026b} proves irrationality measures on three planes of constants. One L\"owner structure produces all three. For all $r,r'\in\Q$, not both zero, \begin{gather*} \mu(rL_{-3}+r'\pi^2)\le20.6622,\qquad\mu(r\zeta(3)+r'\pi^2)\le335.20,\\ \mu(r\zeta(3)+r'\pi^4)\le72.726. \end{gather*} The numbers on these planes were already known to be irrational, by unrelated methods. On the first plane the measure improves the published bound $24781$. On the second it is, as far as we know, the first measure valid on the whole plane. On the third a better bound is known. At special points the bounds are smaller. For example, $\mu(L_{-3})\le13.6839$ and $\mu(L_{-3}\pm\pi^2/9)\le9.4913$. The constructions use differences of values of the trigamma function $\psi'$, \[ \Psi(s,t)=\psi'(s)-\psi'(t)=\sum_{n\ge0}\Bigl(\frac1{(n+s)^2}-\frac1{(n+t)^2}\Bigr)\qquad(0d\rho/\eta$. So $\mu(G)\le d\rho/\eta$ (Proposition~\ref{prop:criterion} gives the precise form). The three numbers come from different places. The degree $d$ comes from the construction (Lemma~\ref{lem:degree}). The neighbourhood exponent $\rho$ comes from a quadrature formula for the coefficient of $G$ (Section~\ref{sec:real}). The difference $\eta$ is a small difference of two large numbers, $D\approx6447$ and $\Lambda\approx6393$ (Figure~\ref{fig:balance}). For OpenAI's construction with our parameters, $D$ and $\Lambda$ differ by about $8$. We lower $\Lambda$ by $22.06$ through the corank of a matrix modulo $p$ (Section~\ref{sec:content}), and raise $D$ by $24.64$ through a bound for a determinant of row functions (Section~\ref{sec:signed}). The result is $\eta\approx54.69$, which is what makes $d\rho/\eta$ as small as $270$. \begin{figure}[ht] \centering \setlength{\unitlength}{0.15cm} \begin{picture}(84,24)(-16,-6) \put(0,-1){\line(1,0){70}} \multiput(0,-1)(10,0){8}{\line(0,-1){0.8}} \put(0,-3.5){\makebox(0,0){$6380$}} \put(20,-3.5){\makebox(0,0){$6400$}} \put(40,-3.5){\makebox(0,0){$6420$}} \put(60,-3.5){\makebox(0,0){$6440$}} \qbezier(-1.5,-1.4)(-1,-0.9)(-0.5,-0.4)\qbezier(-0.9,-1.4)(-0.4,-0.9)(0.1,-0.4) \put(0,14){\framebox(42.8,3){}} \put(42.8,14){\framebox(24.6,3){$+24.6$}} \put(-1,15.5){\makebox(0,0)[r]{decay rate $D$}} \put(0,6){\framebox(12.8,3){}} \put(12.8,6){\dashbox{0.5}(22.1,3){$-22.1$}} \put(-1,7.5){\makebox(0,0)[r]{denominators $\Lambda$}} \put(12.8,3){\line(0,1){11}} \put(67.4,3){\line(0,1){11}} \qbezier(12.8,3.6)(40.1,3.6)(67.4,3.6) \qbezier(12.8,3.6)(13.4,3.85)(14.0,4.1)\qbezier(12.8,3.6)(13.4,3.35)(14.0,3.1) \qbezier(67.4,3.6)(66.8,3.85)(66.2,4.1)\qbezier(67.4,3.6)(66.8,3.35)(66.2,3.1) \put(40,1.6){\makebox(0,0){$\eta=D-\Lambda\approx54.7$}} \put(42.8,19.2){\makebox(0,0){$D_0=6422.8$}} \put(34.9,11.2){\makebox(0,0){$\Lambda_0=6414.8$}} \end{picture} \caption{The exponent balance for the parameters $(58,9,3,5,3)$, per unit of $N^2$. The axis is broken at $6380$. OpenAI's estimates give the decay rate $D_0$ and the denominator exponent $\Lambda_0$, which differ by about $8$. The bound of Section~\ref{sec:signed} adds $24.6$ to $D_0$, and the corank bounds of Section~\ref{sec:content} remove $22.1$ from $\Lambda_0$.}\label{fig:balance} \end{figure} \subsection{New ideas}\label{sec:intro-ideas} \begin{enumerate}[label=\arabic*.,wide] \item \emph{Large corank modulo $p$.} Let $\nu_N(p)$ be the corank modulo $p$ of $p^2$ times the matrix of rational parts. Each unit of corank lowers the denominator by a factor $p$ (Proposition~\ref{prop:valuation}). We show that the corank is of order $N$ for the primes in $113$ intervals inside $(\frac{16}7N,\frac{53}2N]$. The kernel vectors are explicit on most of them. They come from antisymmetric combinations of rows, a reflection formula for Chebyshev polynomials modulo $p$, and the relation $(1-X)^p=1-X^p$. Together these lower the denominator exponent by $22.06$. \item \emph{One counting principle.} All these kernels arise in the same way. One takes a space of row combinations and imposes a few linear conditions that make every entry of the combined row vanish modulo $p$. Then the corank is at least the dimension minus the number of conditions (Theorem~\ref{thm:count}). The four closed formulas are instances. On the remaining intervals of primes we apply the principle to every constant subspace at once (Theorem~\ref{thm:everyU}). Only upper bounds for ranks are then needed, and these come from forced intersections of row spaces. No value of a moment enters. \item \emph{A perturbation bound column by column.} For the neighbourhood of $G$ we write the coefficient of $G$ by a quadrature formula with $53N$ nodes. The nodes are Chebyshev nodes of $106$ short intervals, and the Lagrange polynomials are bounded by the Green functions of the intervals. The perturbed determinant is then a sum of determinants in which some columns are replaced by values of the row polynomials at the nodes. We bound the weight of each replaced column and its generalized Vandermonde factor together. A weighted form of Hadamard's inequality with a parameter $\zeta=\frac34$ makes the dependence on the column cancel, and one contour of radius $\frac34$ treats all columns at once. The part of the weight that depends on the node is carried into the energy bound as an external field. \item \emph{Trial sequences that move with the configuration.} The energy bound with prescribed columns uses trial sequences that depend linearly on the fraction of prescribed columns. Their fields are affine in the parameter, so the suprema are bounded by chords between two end points. A finite cover of the range of the fraction by cells then reduces every case to exact polynomial inequalities. \item \emph{A phase inequality for Pick matrices.} OpenAI's integrand contains a determinant $\det\mathsf L$ of size $n$ that the energy method bounds by $e^{O(n\log n)}$. In the Pick matrix of the nodes the positive nodes carry the weight $-5$. We show that $\det\mathsf L$ loses only a factor $499^n$ when this weight is replaced by $0$, as soon as one phase statistic of the nodes is large enough (Theorem~\ref{thm:window}). The proof interpolates the data of a singular Pick matrix by a finite Blaschke product and follows its argument along an arc of the unit circle. \item \emph{A complete log-gas bound.} With weight $0$ the determinant factors through two Gram determinants. Fischer's inequality and a finite Andr\'eief identity bound it by two logarithmic gases, one in the plane and one in the Joukowski coordinate. These are bounded by tangent inequalities against fixed reference measures. This raises the decay rate by $24.64$. Numerical evaluation for small $N$ suggests that this is most of the true size of $\det\mathsf L$ at the extremal configuration. \item \emph{Near-extremal configurations.} The phase statistic need not be large on every configuration. We show that it is large whenever the configuration comes close to the maximum of OpenAI's energy. On all other configurations OpenAI's bound already suffices. \end{enumerate} We use the parameters $(58,9,3,5,3)$. The general corank theorems of Section~\ref{sec:content} hold for all parameters, and Theorem~\ref{thm:window} does not involve them. \subsection{The constants} Table~\ref{tab:constants} lists the constants for the parameters $(58,9,3,5,3)$, per unit of $N$ or $N^2$. \begin{table}[ht]\centering \begin{tabular}{@{}llr@{}}\toprule degree & $d$ & $53$\\ cost of a prescribed column & $c=c_2+\sigma$ & $3.93$\\ weight of a column & $\psi$ & $50.11$\\ neighbourhood exponent & $\rho=58c+\psi$ & $278.37$\\ decay rate (OpenAI's estimate) & $D_0$ & $6422.83$\\ decay rate (with the bound for $\det\mathsf L$) & $D$ & $6447.47$\\ denominator exponent (with the corank gain) & $\Lambda$ & $6392.78$\\ difference & $\eta=D-\Lambda$ & $54.68$\\ \midrule bound & $d\rho/\eta$ & $269.78\ldots$\\\bottomrule \end{tabular} \caption{The constants of the proof, truncated. The formal proof uses exact rational numbers.} \label{tab:constants}\end{table} Without the corank gain and the bound for $\det\mathsf L$, $\eta$ would be about $8$, and the method would give an exponent above $1800$. \subsection{Organization} Section~\ref{sec:overview} recalls OpenAI's construction for general parameters and proves the criterion that turns the three properties of \S\ref{sec:intro-need} into a measure. Section~\ref{sec:setup} bounds the denominators and proves nonvanishing at every rational point. Section~\ref{sec:content} proves the corank bounds, which give the arithmetic gain. Sections~\ref{sec:real}--\ref{sec:signed} treat the analytic side. Section~\ref{sec:real} extends OpenAI's estimate to a neighbourhood of $G$. Section~\ref{sec:energy} recalls the energy method with prescribed columns, and Section~\ref{sec:signed} proves the bound for the row determinant. Section~\ref{sec:proof} assembles the proof, and Section~\ref{sec:remarks} discusses the method. The certified numerical inequalities are collected in the appendices, and Section~\ref{sec:formal} describes the formalization. \subsection{Notation} We use the notation of OpenAI's paper \cite{OpenAI2026} for the construction and refer to it for details. Citations of the form \cite[Prop.~2.1]{OpenAI2026} point to its numbered statements. We write $\log$ for the natural logarithm, $v_p$ for the $p$-adic valuation and $\vartheta(x)=\sum_{p\le x}\log p$ for Chebyshev's function. $T_k$ and $U_k$ are the Chebyshev polynomials of the first and second kind, with $U_{-1}=0$, $T_{-k}=T_k$ and $U_{-k-1}=-U_{k-1}$. We write $[X^k]F$ or $F[k]$ for the coefficient of $X^k$ in a polynomial $F$, $y_+=\max(y,0)$, and $\operatorname{rev}_EF(X)=X^EF(1/X)$ for the reversal of a polynomial of degree at most $E$. \section{The construction and the criterion}\label{sec:overview} \subsection{Designs} OpenAI's determinant depends on five integer parameters. Following Carreira \cite{Carreira2026} we call a quintuple $(n_0,b_0,q_0,g_0,h_0)$ of integers with $n_0,b_0,q_0,h_0\ge1$, $g_0\ge0$ and $2g_00\}}\frac{f(t)}{t}D_r(t),\qquad \varphi_k(s)=s^{b+k}(1-s)^q . \] The determinant studied by OpenAI is that of the $n\times n$ matrix \begin{equation}\label{eq:matrix} \cA_N(r,k)=\int_{-1}^1\int_0^1 A_r(t)\,\frac{\varphi_k(s)}{1-ts}\,ds\,dm(t) \qquad(0\le r,k0$ satisfy \begin{gather*} \omega<\lambda,\quad \delta<\tfrac13,\quad d\le(\lambda-\omega)\bigl(\eta+2d(1-3\delta)\bigr),\\ 1<2\bigl(\omega-\delta(2\lambda-\omega)\bigr),\quad \eta<2d,\quad d+4d\delta\lambda<\eta\lambda . \end{gather*} For each prime $p$ let $Q_p\in\Z[X]$, and suppose that the following hold for all large $p$. \begin{enumerate}[label=(\alph*)] \item $\deg Q_p\le dp$. \item $Q_p(a/b)\ne0$ whenever $a\in\Z$, $b\ge1$ and $p^2\nmid b$. \item For all primes $pb^{-\rho\lambda}$ for all integers $a$ and all sufficiently large $b$. \end{proposition} \begin{proof} Let $a/b$ be a rational number with $b$ large. Put $u_1=\lambda\log b$ and $u_0=(\lambda-\omega)\log b$, and suppose $|\xi-a/b|\le b^{-\rho\lambda}$. By the prime number theorem, $(1-\delta)t\le\vartheta(t)\le(1+\delta)t$ for all large $t$, and $u_0\to\infty$ with $b$. If every prime $p\in(u_0,u_1]$ had $p^2\mid b$, then $2(\vartheta(u_1)-\vartheta(u_0))\le\log b$. But \begin{align*} 2\bigl(\vartheta(u_1)-\vartheta(u_0)\bigr)&\ge2\log b\,\bigl((1-\delta)\lambda-(1+\delta)(\lambda-\omega)\bigr)\\ &=2\log b\,\bigl(\omega-\delta(2\lambda-\omega)\bigr)>\log b \end{align*} by the fourth hypothesis. So there are primes in $(u_0,u_1]$ whose squares do not divide $b$. Let $p$ be the \emph{largest} one. Then every prime $p'\in(p,u_1]$ has $p'^2\mid b$. Let $T\le\min(u_1,2p)$ with $T\ge p$, and let $\Pi_T$ be the product of the $p'^2$ over the primes $p'\in(p,T]$. Then $\Pi_T\mid b$ and $\log\Pi_T=2(\vartheta(T)-\vartheta(p))$. Consider the integer $b^{dp}Q_p(a/b)=\sum_ic_ia^ib^{dp-i}$. Each of its terms is divisible by $p'^{2dp}$ for each of these $p'$. Indeed $p'^{2i}$ divides $c_i$ by (c), and $p'^2\mid b$ gives $p'^{2(dp-i)}\mid b^{dp-i}$. So $\Pi_T^{dp}$ divides the integer. Since $p^2\nmid b$, $Q_p(a/b)\ne0$ by (b). So $b^{dp}Q_p(a/b)/\Pi_T^{dp}$ is a nonzero integer, and \begin{equation}\label{eq:recovered} 1\le\Bigl(\frac b{\Pi_T}\Bigr)^{dp}\,|Q_p(a/b)| . \end{equation} Since $\rho p\le\rho u_1=\rho\lambda\log b$, we have $|a/b-\xi|\le e^{-\rho p}$. So $|Q_p(a/b)|\le e^{-\eta p^2}$, and \eqref{eq:recovered} gives \begin{equation}\label{eq:recovered2} \eta p\le d\bigl(\log b-2(\vartheta(T)-\vartheta(p))\bigr). \end{equation} Suppose first that $u_1\le2p$, and take $T=u_1$. With $\vartheta(u_1)-\vartheta(p)\ge u_1-p-2\delta u_1$, \eqref{eq:recovered2} reads \[ \eta p\le d\bigl(\log b-2u_1+2p+4\delta u_1\bigr). \] Hence $(\eta-2d)p\le d\log b-2du_1+4d\delta u_1$. As $\eta<2d$ and $p\le u_1$, the left side is at least $(\eta-2d)u_1$. So $\eta u_1\le d\log b+4d\delta u_1$, that is, $\eta\lambda\le d+4d\delta\lambda$. This contradicts the last hypothesis. Suppose now that $2pu_0=(\lambda-\omega)\log b$, this gives $(\lambda-\omega)\bigl(\eta+2d(1-3\delta)\bigr)2p$, and then (c) recovers only the primes up to $2p$. The third hypothesis treats this case. A small prime $p$ makes the decay $e^{-\eta p^2}$ weak, but it also makes the height $b^{dp}$ small, and the primes of $(p,2p]$ alone already win. So $\lambda$ need not exceed $2\omega$, and the bound $\rho\lambda$ may lie below $\rho$. For Catalan's constant we take $d=53$, $\omega=\frac35$, $\delta=2.9\cdot10^{-7}$ and $\lambda=270/\rho$. The last hypothesis then reads $270>\mu_*$ with $\mu_*=d\rho/(\eta-4d\delta)=\Mustar\ldots$ The formal proof checks the six hypotheses with $\rho\lambda$ in place of $270$ for every $\rho\lambda\in[269.787,400]$. \section{Denominators and nonvanishing}\label{sec:setup} This section proves the arithmetic properties of $\Delta_N$: the denominators of its coefficients and its nonvanishing at rational points. Carreira \cite{Carreira2026} checked that OpenAI's argument goes through for parameters satisfying a few conditions. We need more. The entries must be controlled modulo $p^2$, since the corank bounds of Section~\ref{sec:content} work there. And nonvanishing is needed at every rational point $a/b$ with $p^2\nmid b$, not only at $G$. \subsection{Corank and valuation} We use repeatedly the following elementary fact. \begin{lemma}\label{lem:rank-valuation} Let $M$ be an $n\times n$ matrix with entries in $\Z_{(p)}$ whose reduction modulo $p$ has rank at most $n-e$. Then $v_p(\det M)\ge e$. \end{lemma} \begin{proof} Factor the reduction through $\F_p^{n-e}$ and lift the two factors to integral matrices $U,V$, so that $M=UV+pE$ with $E$ integral at $p$. Expand the determinant multilinearly in the columns. Every term that uses more than $n-e$ columns of $UV$ vanishes, and every other term has at least $e$ factors $p$. \end{proof} \subsection{The digit formula} The reduction of $p^2\cA^0_N$ modulo $p$ is described by the base-$p$ digits of the coefficients of the row polynomials. This is the one place where the arithmetic of the moments enters, and it does not depend on the parameters. For an odd prime $p$ and $c\in\F_p^n$ put \[ P_c=\sum_rc_rP_r,\qquad D_c=\sum_rc_rD_r,\qquad W=(1-X^2)^{(p-1)/2}, \] reading the $P_r,D_r$ as polynomials in $X$. \begin{lemma}[digit formula]\label{lem:digits} Let $p$ be an odd prime with $HH$ the matrix $\cA^0_N+z\cL_N$ is $p$-integral. Its integral over $[1,78]$ is $24269/4$, and $f=116=2n_0$ on $[0,1]$. Since $f(\theta)=2n_0$ for $\theta\le59/2$, every prime of the range $(\frac{16}7N,\frac{53}2N]$ of the corank bounds of Section~\ref{sec:content} has $\ell_N(p)=2n$. There Proposition~\ref{prop:valuation} improves the bound $-2n$ to $\nu_N(p)-2n$. The primes $p\le\sqrt H$ are not covered by these bounds. For them we use the cruder bound $6n\log_pH$ for the exponent, which contributes $O(\sqrt N\,N\log N)=o(N^2)$ in total. \subsection{The prime $2$} At $p=2$ the entries are not integral, and the bound comes from the expansion of the determinant into minors, as in \cite[\S3]{OpenAI2026}. Its exponent is governed by a quadratic form in the numbers $m,\ell$ of columns of two exceptional types, \[ \tfrac32m^2+\tfrac12\ell^2-(H+C-b)\,m-C\,\ell+Cn . \] The following identity shows that it is at least $-\Lambda_2$ whenever $m,\ell\ge0$, $m+\ell\le n$ and $C\ge n$, where \begin{equation}\label{eq:A2} \Lambda_2=\frac{(n+H-b)^2}8-\frac{n^2}2 . \end{equation} With $s=n-m-\ell$ and $m_*=(n+H-b)/4$, \[ \tfrac32m^2+\tfrac12\ell^2-(H+C-b)m-C\ell+Cn+\Lambda_2 =2(m-m_*)^2+s\,(C-n+m)+\tfrac12s^2\ge0 . \] This gives a $2$-adic bound $v_2(\Delta_N(i))\ge-\Lambda_2-O(N\log N)$, uniform in the integer $i$. For our parameters $\Lambda_2=(58+78-9)^2/8-58^2/2=2673/8$ in units of $N^2$. \subsection{The common denominator} The corank bounds of Section~\ref{sec:content} give a function $\kappa$ of $\theta=p/N$ (Theorem~\ref{thm:profile-D}). We use it here as an input. \begin{proposition}\label{prop:denominators} For the parameters $(58,9,3,5,3)$ there are positive integers $\cD_N$ with the following properties. $\cD_N\Delta_N(i)\in\Z$ for every integer $i$, and $p'^{2n}\mid\cD_N$ for every prime $N0$ and all large $N$ \begin{gather*} \log\cD_N\le(\Lambda+\varepsilon)N^2,\\ \Lambda=\frac{2673}8\log2+116+\frac{24269}4-\int_0^\infty\kappa(\theta)\,d\theta=\LamDen\ldots \end{gather*} \end{proposition} Here $\frac{2673}8\log2$ comes from the prime $2$, and $116+\frac{24269}4$ is the area of $f$. The last term is the gain from the corank. \begin{proof} For primes $p\le H$ with $HH$ the matrix is $p$-integral. So $\cD_N\Delta_N(i)\in\Z$. A prime $p'\in(N,2N]$ has $\nu^*_N(p')=0$, since no interval of Theorem~\ref{thm:profile-D} contains it, and $\ell_N(p')=2n$. So $p'^{2n}\mid\cD_N$. Here $N\ge79$ guarantees $p'>\sqrt H$. For the size of $\cD_N$ we use the prime number theorem in the weighted form \begin{equation}\label{eq:weighted-pnt} \frac1N\sum_{p}F(p/N)\log p\longrightarrow\int_0^\infty F(\theta)\,d\theta\qquad(N\to\infty), \end{equation} for bounded, Riemann integrable, compactly supported $F$. By \eqref{eq:weighted-pnt} the sum of $\ell_N(p)\log p$ over the odd primes is $N^2\int_0^{78}f+o(N^2)$, with $\int_0^{78}f=116+24269/4$. The primes in the intervals of Theorem~\ref{thm:profile-D} lie in $(16N/7,53N/2]$ and have $\ell_N(p)=116N$. So their exponent is at most $116N-N\kappa(p/N)+2$. The sum of $(N\kappa(p/N)-2)\log p$ over these primes is $N^2\int\kappa+o(N^2)$, again by \eqref{eq:weighted-pnt}. The $2$-adic part contributes $\Lambda_2\log2\,N^2+o(N^2)$. \end{proof} The polynomial $\Delta_N$ has degree at most $n-g=d_0N$ and takes values in $\cD_N^{-1}\Z$ at the integers. An integer-valued polynomial of degree $m$ has coefficients in $\frac1{m!}\Z$, by Newton's interpolation formula. So \begin{equation}\label{eq:QN} Q_N=(d_0N)!\,\cD_N\,\Delta_N(X)\in\Z[X],\qquad \deg Q_N\le d_0N, \end{equation} and $\log(d_0N)!=o(N^2)$. In this way the gain from the corank enters through the integer polynomial itself, not through a separate common factor of its coefficients. The criterion needs the squares of the primes in $(N,2N]$ to divide the coefficients of $Q_N$ in a graded way. \begin{lemma}\label{lem:squares} For all large $N$ and every prime $N4H_0$, with $H_0=H/N$, so that $H=H_0pe$. The block consists of the columns $k(J,\varsigma)=Jp+\varsigma-(b+q)$ for $1\le J\le5$ and $\varsigma\in \mathcal M$. These indices are nonnegative by the hypothesis $b+q\le p+\varsigma_0$, and smaller than $n$ because $5p+\varsigma_1-1=L-1$. There are $5w$ of them. As $\varsigma\deg Y_c$. So, if $Q_c[up+\varsigma']=0$ for $u=2,\dots,5$ and all $\varsigma'\in \mathcal M$, then \[ \sum_{t=0}^e(-1)^t\binom et\,Q_c[6p+\varsigma-t]=0\qquad(\varsigma_0+e\le\varsigma<\varsigma_1). \] This is a recurrence with unit leading coefficient. So if $Q_c[6p+\varsigma]=0$ for the first $e$ values $\varsigma=\varsigma_0,\dots,\varsigma_0+e-1$, then it vanishes for all $\varsigma\in \mathcal M$. \emph{Counting.} Impose on $c$ that the values at the $n-5w$ columns outside the block vanish, that $Q_c[up+\varsigma]=0$ for $2\le u\le5$ and $\varsigma\in \mathcal M$ ($4w$ conditions), and that $Q_c[6p+\varsigma]=0$ for the first $e$ values of $\varsigma$. By the three steps above all the values at the block columns vanish. These are at most $n-5w+4w+e=n-(w-e)$ linear conditions, so the left kernel has dimension at least $w-e$. Finally, $w-e$ equals $\varsigma_1-q-e=n+b+q+h-(11p+1)/2$ if $\varsigma_0=q$, and $(p-1)/2-g$ if $\varsigma_0=H+q-6p$. \end{proof} \subsection{The digits $1$ to $4$} \begin{theorem}\label{thm:S} Let $e=(p+1)/2-q-h\ge0$ and $n\le4p+2g+1$, and put \[ m_0=\max(b+q,\,L+\chi-4p),\qquad m_1=\min(p,\,L-3p,\,A+1). \] Then $\nu_N(p)\ge m_1-m_0-e$. \end{theorem} \begin{proof} Let $\mathcal M=[m_0,m_1)$ and $w=m_1-m_0>e$. For $\varsigma\in \mathcal M$ we have $\varsigma\le A$, so the digit $u=0$ of $Q_c$ and $\widehat D_c$ vanishes. The inequality $\varsigma\ge L+\chi-4p$ gives that the digits $u\ge4$ of $\widehat D_c$ and $u\ge5$ of $Q_c$ vanish. Finally $3p+\varsigma10$ and $\theta\ge39/2$, with $k_T=\min(5N,52N-1-2p,75N-3p)$. It gives $3p-(n+q+h)=3\theta-64$ up to $\theta=23$ and the bound $g=5N$ up to $\theta=70/3$. From $\theta=70/3$ to $\theta=26$ the hypotheses of Theorem~\ref{thm:U} hold. They read $3p\ge L$, $3p\le H$, $2p\le n+q$ and $2h\le p+1$. The bound is $\min(n,2p+2g+1)-g-2p-1$, which is $5N$ for $\theta\le24$ and $53N-2p-1$ above. In each case the value of the theorem is the value in the table minus at most one unit, from the integer rounding. The same kind of check, with linear inequalities in $N$, $p$ and the parameters, gives the bounds of Section~\ref{ssec:families} on their intervals, and Appendix~\ref{app:windows} gives the others. This proves Theorem~\ref{thm:profile-D}. The loss of $2$ in its statement covers the integer rounding of all the bounds. \section{The neighbourhood of $G$}\label{sec:real} This section and the next two prove that $\Delta_N$ is small on a whole neighbourhood of $G$. The idea is to write the perturbation $\Delta_N(x)-\Delta_N(G)$ as a sum of determinants in which some columns of $\cA_N$ are replaced by values of the row functions at prescribed nodes. Each such determinant is bounded by OpenAI's energy method, extended to configurations with prescribed points (Section~\ref{sec:energy}). The loss caused by each replaced column, in the exponent, determines $\rho$. OpenAI's method has two cases. In the first it already gives the improved decay rate. In the second it falls short when few columns are replaced, and Section~\ref{sec:signed} closes that gap. \subsection{The statement} OpenAI's energy method bounds $|\Delta_N(G)|$ by $\exp\bigl(-(n_0^2(U-\frac12\log2)+o(1))N^2\bigr)$, where $U$ is an upper bound for a discrete logarithmic energy (Section~\ref{sec:energy}). For our parameters we certify $U=\Utarget$, so OpenAI's estimate gives the decay rate $D_0=n_0^2(U-\frac12\log2)=\DecayOA\ldots$ Let \begin{gather*} \Gamma_0=\Gain,\qquad U_0=U+\Gamma_0/n_0^2=\Uzero\ldots,\\ D=n_0^2(U_0-\tfrac12\log2)=\Decay\ldots \end{gather*} The number $\Gamma_0$ is the improvement of the decay rate proved in Section~\ref{sec:signed}. Put \begin{equation}\label{eq:rho-def} \rho=n_0c+\psi,\qquad c=c_2+\sigma, \end{equation} with $c_2=\Ctwo$, $\sigma=\Sigm$ and $\psi=\Psiw$, so that $\rho=\Rho$. Here $c$ is the increase of the energy bound per prescribed column (Section~\ref{sec:energy}), and $\psi$ bounds the logarithmic size of the weight and the Vandermonde factor of one column, per unit of $N$ (\S\ref{ssec:weights}). \begin{theorem}\label{thm:real} For our parameters, uniformly for real $x$ with $|x-G|\le e^{-\rho N}$, \[ |\Delta_N(x)|\le e^{-(D+o(1))N^2}\qquad(N\to\infty). \] \end{theorem} We use the following abbreviations for general parameters: \begin{gather*} \alpha_A=\frac An,\quad \beta_h=\frac{2h_0}{n_0},\quad \gamma_H=\frac Hn,\quad \gamma_d=\frac{d_0}{n_0},\\ a_0=1-\frac{b_0+q_0+g_0}{n_0},\quad a_1=3-\frac{b_0+q_0+2g_0}{n_0}. \end{gather*} For our parameters these are $\alpha_A=11/29$, $\beta_h=3/29$, $\gamma_H=39/29$, $\gamma_d=53/58$, $a_0=41/58$ and $a_1=76/29$. \subsection{A quadrature formula for the coefficient of $G$}\label{ssec:quadrature} Let $\varepsilon=x-G$. Then $\Delta_N(x)=\det(\cA_N+\varepsilon\cL_N)$. The question is how large the columns of $\cL_N$ are compared with those of $\cA_N$. By \eqref{eq:L} the entries of $\cL_N$ are linear functionals of the coefficients of the row polynomials. These polynomials have bounded degree, so the functionals can be written exactly as finite combinations of point evaluations. Let $t_1,\ldots,t_J$ be distinct points, with $J$ at least the number of degrees of freedom of the rows. Lagrange interpolation gives \begin{equation}\label{eq:quad} \cL_N(r,k)=\sum_{a=1}^JP_r(t_a)\,\beta_{a,k}, \end{equation} where the weight $\beta_{a,k}$ is the value of the same functional on the $a$-th Lagrange basis polynomial $\ell_a$. This is a quadrature formula of Gauss type. It makes the columns of $\cL_N$ combinations of point values of the polynomials $P_r$. So the perturbed determinant can be treated by the energy method. The cost is the size of the weights. We take $J=d_0N=53N$ nodes, which are Chebyshev nodes of short intervals. Let $[a_c,b_c]$, $1\le c\le106$, be fixed intervals with dyadic end points, $48$ in $[-1,0)$ and $58$ in $(0,1)$, adjacent within each side. Their union is $[-0.99931266\ldots,-0.37573294\ldots]\cup[0.19559072\ldots,0.99999983\ldots]$. In the coordinate $x$ with $t=2x/(1+x^2)$ it is contained in \[ E_x=[-\tfrac{2409}{2500},-\tfrac{19501}{100000}]\cup[\tfrac{98749}{1000000},1]. \] So the nodes avoid a neighbourhood of $x=0$, where the fields of the energy method are largest. We pair the intervals $(2i-1,2i)$ within each side, which gives $24+29=53$ pairs. The two members of a pair receive $K_c=\lceil N/2\rceil$ and $\lfloor N/2\rfloor$ nodes, so that $\sum_cK_c=53N$. Let $\ell_c(z)=(2z-a_c-b_c)/(b_c-a_c)$ and $w_c=b_c-a_c$. The nodes in $[a_c,b_c]$ are the zeros of $T_{K_c}\circ\ell_c$. They lie in the open interval, so all nodes are distinct. Let $G(w)=\log|w+\sqrt{w^2-1}|$, with the branch of modulus at least $1$, be the Green function of $\C\setminus[-1,1]$. Put $G_c=G\circ\ell_c$ and \begin{equation}\label{eq:Ldef} L(z)=\tfrac12\sum_c\bigl(G_c(z)+\log(w_c/4)\bigr). \end{equation} \begin{lemma}\label{lem:lagrange} There is a constant $C$ such that for all $N$, every node $t_a$ and every $z\in\C$ with $|z|\le1$, the Lagrange polynomial $\ell_a$ of $t_a$ satisfies $|\ell_a(z)|\le CN\,e^{N(L(z)-L(t_a))}$. \end{lemma} \begin{proof} We use three facts about $T_K$. First, $|T_K(w)|\le e^{KG(w)}$, since $2T_K(w)=\phi^K+\phi^{-K}$ with $\phi=w+\sqrt{w^2-1}$. Second, $|T_K(v)|=\cosh(KG(v))\ge\frac12e^{KG(v)}$ for real $|v|\ge1$. Third, the Lagrange polynomial $\lambda_j$ of the $j$-th of the $K$ Chebyshev nodes satisfies $|\lambda_j(w)|\le Ke^{(K-1)G(w)}$ for all $w\in\C$. For the third, write $w=\frac12(u+u^{-1})$ with $|u|=e^{G(w)}\ge1$ and the node as $\frac12(v+v^{-1})$ with $|v|=1$. Then \[ \frac{T_K(w)}{w-\frac12(v+v^{-1})}=\Bigl(\sum_{iC_0+D_\Gamma-48\log\tfrac34=50.1113394\ldots \] We take $\psi=\Psiw$, which leaves a strict margin for the terms $o(1)$. The factor $e^{N\hat\delta(t_a)}$ depends on the node. We do not bound it here. It is carried into the energy bound, where it raises the field of each prescribed point $y$ by $\hat\delta(\mathsf t(y))/58$, with $\mathsf t(y)=2y/(1+y^2)$. The remaining part of $\rho$ is $n_0c$, the increase of the energy bound caused by the prescribed columns. It comes from Section~\ref{sec:energy}, and with it $\rho=58\cdot\Cost+\Psiw=\Rho$. \section{The energy method with prescribed columns}\label{sec:energy} This section recalls OpenAI's energy method, extends it to determinants with prescribed columns, and proves Theorem~\ref{thm:real} assuming the bound of Section~\ref{sec:signed}. Throughout, $\alpha_1=\AlphaOne$. \subsection{OpenAI's energy}\label{ssec:energy} By Andr\'eief's identity, $\det\cA_N$ is an integral over $n$ points $x_i\in(-1,1)$ and $n$ points $s_j\in(0,1)$. OpenAI bounds the integrand by $e^{n^2E}$, where $E$ is a discrete logarithmic energy of the points, written in the coordinate $x$ with $t=2x/(1+x^2)$ \cite[\S5]{OpenAI2026}. It has the form \begin{multline*} n^2E=\sum_iV_1(x_i)+\sum_jV_2(s_j)+\sum_{i0$ on $(0,1]$, the value with $\alpha g$ is at most the value with $\gamma g$ for every $\gamma\ge\alpha$. The factor $e^{N\hat\delta}$ of the weights adds $\hat\delta(\mathsf t(x))/58$ to the field of every prescribed point $x$, where $\mathsf t(x)=2x/(1+x^2)$. Indeed $n^2\cdot\frac1n\sum_j\hat\delta(\mathsf t(y_j))/58=N\sum_j\hat\delta(\mathsf t(y_j))$. Every step of OpenAI's argument remains valid uniformly in the number of free points. The regularization of the logarithmic interactions, the diagonal corrections and the completion of squares only involve sums over the points present. For the prescribed points of the first kind the relevant field is \[ F_1(x)=-\gamma_d\log(1+x^2)-2\mathsf T(u,x)+\tfrac1{58}\hat\delta(\mathsf t(x)),\qquad x\in E_W . \] For those of the second kind it is \[ F_2(y)=-\gamma_d\log(1+y^2)+\frac{n_0-2g_0-1}{n_0}\log y-u_1y+\tfrac1{58}\hat\delta(\mathsf t(y)) \] for $y\in E_W$ with $0\Kcoef), \end{equation} with $\Kcoef$ Chebyshev coefficients and finitely many geometric terms. The geometric terms have a simple meaning. Since $\sum_{k\ge1}z^kT_k(x)/k=-\frac12\log(1-2xz+z^2)$, a term $(z,r)$ contributes $-\frac12\operatorname{Re}\,r\log(1-2xz+z^2)$ to the potential $\mathsf T(u,x)$. If $x_0+i\epsilon_0=\frac12(z+z^{-1})$, then $1-2xz+z^2=2z(x_0+i\epsilon_0-x)$. So the term is, up to a constant, the logarithmic potential of a point charge of complex strength $r$ at $x_0+i\epsilon_0$ and its conjugate. The extremal measure has a gap in its support around $x=0$. Its potential has a logarithmic singularity at $x=0$ and square-root behaviour at the edges of the gap. A Chebyshev expansion converges slowly near such points, while a charge at a small distance $\epsilon_0$ reproduces the local behaviour with one term. This is also why the quadrature nodes of Section~\ref{sec:real} avoid a neighbourhood of $x=0$. \emph{Moving trial sequences.} The best trial sequence depends on $\alpha$. We use families \[ u(\tau)=u+\tau\,\mathrm du,\qquad v(\tau)=v+\tau\,\mathrm dv\qquad(0\le\tau\le1), \] with fixed finitely supported increments and with $\tau=\tau(\alpha)$ an explicit piecewise linear function. The fields $X$, $Y$, $F_1$ and $F_2$ are affine in $\tau$. So the supremum of a field at $\tau$ is at most the chord between the suprema at $\tau=0$ and $\tau=1$, or, for the kept base fields, the supremum at $\tau=0$ plus $\tau$ times the supremum of the increment. The norms are quadratic in $\tau$. All bounds then become polynomials in $\alpha$ on each piece of $\tau(\alpha)$. Appendix~\ref{app:field} gives the charges, the increments and the certified bounds. \subsection{The two cases at the improved rate}\label{ssec:cases} Theorem~\ref{thm:real} asks for the rate $U_0=U+\Gamma_0/n_0^2$ instead of $U$, at the slope $c$. In both cases this reduces to finitely many polynomial inequalities in $\alpha$, each of the form $V(\alpha)\le\sigma\alpha$ on a closed interval with polynomial caps. Each is proved by a cover of the interval by subintervals on which the leading term dominates, in exact rational arithmetic. The bound at an interior $\alpha$ of a cell uses the caps at the right end point of the cell, since $g>0$. A finite cover by cells and the uniform continuity of the caps then give the bound for all $\alpha$, with an error that tends to zero. In the constrained case we use the increment with $24$ Chebyshev coefficients and $\tau=\min(1,\frac{50}{11}\alpha)$. The inequality holds on $[0,1]$, which is covered by the $6$ pieces with end points $0,\frac{11}{50},\frac3{10},\frac25,\frac35,\frac45,1$. In the case with doubled interaction we use the increment with $40$ Chebyshev coefficients and $\tau=1$. The inequality of Lemma~\ref{lem:energy-ineq} holds on $[\alpha_1,1]$, covered by the $6$ pieces with end points $\alpha_1,\frac15,\frac3{10},\frac25,\frac35,\frac45,1$. For $\alpha<\alpha_1$ it fails, because the extra term $U_0-U$ is absorbed by the term $c\alpha$ only once $\alpha$ is not small. So a new argument is needed only in the case with doubled interaction with fewer than $\alpha_1n$ prescribed columns. Section~\ref{sec:signed} supplies it, with the same slope $c$. Precisely, it proves the following. \begin{proposition}\label{prop:signed-goal} In the case with doubled interaction, for every configuration with $m\le\alpha_1n$ prescribed columns at the nodes $t_{a_1},\ldots,t_{a_m}$, the integrand of Lemma~\ref{lem:andreief}, with the generalized Vandermonde determinant replaced by the bound of Lemma~\ref{lem:binomvand}, is at most \[ \exp\Bigl(n^2\bigl(-U_0+\tfrac12\log2+c\alpha+o(1)\bigr)-N\sum_j\hat\delta(t_{a_j})\Bigr)\,C_*^m\prod_j\mathcal B_\zeta(k_j),\qquad \alpha=m/n, \] with an absolute constant $C_*$. \end{proposition} \begin{proof}[Proof of Theorem~\ref{thm:real}, given Proposition~\ref{prop:signed-goal}] By Proposition~\ref{prop:signed-goal} and the discussion above, in both cases a determinant $M_f$ with $m$ prescribed columns $k_1,\ldots,k_m$ at the nodes $t_{a_j}$ is at most \[ C_*^m\prod_j\mathcal B_\zeta(k_j)\,e^{n_0cNm-N\sum_j\hat\delta(t_{a_j})}\exp\bigl(n_0^2N^2(-U_0+\tfrac12\log2+o(1))\bigr). \] The factor $C_*^m$ comes from the expansion of the row functions at the prescribed points. By \eqref{eq:wk} and \S\ref{ssec:weights}, the weight of the column $k_j$ at the node $t_{a_j}$, times $\mathcal B_\zeta(k_j)$, is at most $e^{(\psi+\hat\delta(t_{a_j})+o(1))N}$. The terms $\hat\delta$ cancel. So for every column $k$ and every node $a$ the weight and the factor of $M_f$ that belong to the pair $(k,a)$ have the product \[ \log(B_kC_k)\le\bigl(n_0c+\psi+o(1)\bigr)N=(\rho+o(1))N , \] and Lemma~\ref{lem:expansion} applies with this bound for each pair. For $|\varepsilon|\le e^{-\rho N}$ Lemma~\ref{lem:expansion} gives a factor $(1+O(N^5))^n=e^{o(N^2)}$. With $D=n_0^2(U_0-\frac12\log2)$ this is the claim. \end{proof} \section{A bound for the row determinant}\label{sec:signed} This section proves Proposition~\ref{prop:signed-goal}, which completes the proof of Theorem~\ref{thm:real}. In the case with doubled interaction OpenAI's integrand contains a determinant $\det\mathsf L$ of size $n$ that the energy method bounds by $e^{O(n\log n)}$. We replace this determinant by a simpler one, with a loss of only $e^{O(N)}$, and we bound the simpler one by a Gram determinant. The argument has three steps. The first step is a comparison. Let $F(w)$ be the Pick matrix of the nodes with weight $w$ on the positive nodes. Then $\det\mathsf L$ is $\det F(-5)$ up to a known factor, and a positivity argument bounds it by $499^n\det F(0)$ as soon as $F(-\frac1{100})\ge0$ (\S\ref{ssec:rowdet}). The second step shows that $F(-\frac1{100})\ge0$ whenever a phase inequality holds for the nodes (\S\ref{ssec:window}, \S\ref{ssec:windowproof}), and that this inequality holds on every configuration for which the energy bound is not already sufficient (\S\ref{ssec:alternative}). The third step bounds $\det F(0)$, which is a Vandermonde determinant of a Laurent system. It splits into a Vandermonde determinant of the positive nodes and a determinant of folded functions at the negative nodes. Fischer's inequality and Andr\'eief's identity turn the latter into two log-gas integrals (\S\ref{ssec:gram}). Reference measures then bound these integrals, and a numerical inequality closes the proof (\S\ref{ssec:budget}). \subsection{The row determinant and the comparison}\label{ssec:rowdet} In the case with doubled interaction OpenAI's integrand contains the determinant of the row functions at points $z_1,\ldots,z_n\in(-1,1)\setminus\{0\}$, which we call the nodes \cite[Prop.~5.2]{OpenAI2026}. They are the free points $x_i$ and the prescribed points $y_j$ together. We write $\det\varphi(z)$ for this determinant. With $\mathsf e=n-1-2g$ it factors as \begin{equation}\label{eq:detphi} \det\varphi(z)=C_{n,g}\,V(z)\prod_a|z_a|^{-(n-1-g)}\det\mathsf L(z), \end{equation} where $C_{n,g}$ depends only on $n$, $g$ and the number of nodes of each kind, and $V(z)=\prod_{a0$, all eigenvalues of $Y_T^{-1/2}Z_TY_T^{-1/2}$ have modulus at most $499$, which is the claim. The singular case follows by replacing $Y_T$ by $Y_T+tI$ and letting $t\to0$. The second statement follows from \eqref{eq:minors}. \end{proof} So everything reduces to two questions: when is $F(-\frac1{100})\ge0$, and how large is $\det\Xi^0_S$? \subsection{A phase criterion for positivity}\label{ssec:window} The matrix $F(-c)$ is a Pick matrix with a small negative weight on the positive nodes. The following theorem gives a criterion for its positivity that depends only on the positions of the positive nodes. It holds for all configurations and has no relation to our parameters. \begin{theorem}\label{thm:window} Let $z_1,\ldots,z_n\in(-1,1)\setminus\{0\}$ be distinct, with $k_+$ positive and $k_-$ negative entries. Let $t\ge0$ be an integer, let $c=\frac1{100}$, and let $D_a=B'(z_a)z_a^t$. For a positive node $p$ put $\rho_p=(1-p)/(1+p)$. If \begin{equation}\label{eq:window} t\arctan4-\sum_{p>0}\arctan(4\rho_p)>\frac{\pi}{2}(k_-+1)+\frac{17}4\,\frac{c^2}{1-c^2}\,k_+, \end{equation} then $\cK+\operatorname{diag}(E_-,-cE_+)\ge0$. \end{theorem} The number $\arctan4$ is the half-angle of the arc of the unit circle around $z=-1$ with end point $z_0=(4i-1)/(4i+1)$. The proof compares two phases on this arc. We apply the theorem with $t=\mathsf e$. \subsection{Proof of Theorem~\ref{thm:window}}\label{ssec:windowproof} Let $f_a(\gamma)=\gamma z_a^t$ at a negative node and $f_a(\gamma)=-cz_a^t$ at a positive node. Let $W(\gamma)_{ab}=(1-f_a(\gamma)f_b(\gamma))\cK_{ab}$ be the Pick matrix of these data. By Lemma~\ref{lem:pick} it suffices to prove $W(1)\ge0$. With $X=\operatorname{diag}(z_a^t)$ we have \[ W(-c)=(1-c^2)\cK+c^2(\cK-X\cK X), \] and $\cK-X\cK X\ge0$ as in Lemma~\ref{lem:pick}. So $W(-c)>0$. If $W(1)$ is not positive semidefinite, then there is $\gamma\in(-c,1)$ with $W(\gamma)\ge0$ singular. Fix such a $\gamma$ and write $W=W(\gamma)$, $f_a=f_a(\gamma)$. Note that $|f_a|<1$. Let $\mathcal T$ be a minimal set of nodes such that the principal submatrix $W_{\mathcal T}$ is singular. Since the diagonal entries of $W$ are positive, $|\mathcal T|\ge2$. By minimality $W_{\mathcal T}$ has a kernel vector $v$ with all entries nonzero, and every proper principal submatrix is positive definite. From now on only the nodes of $\mathcal T$ are used, and $k_+$ and $k_-$ count these nodes. Inequality \eqref{eq:window} still holds, because removing a node lowers the right side or raises the left side. If all nodes of $\mathcal T$ have the same sign, the data are a constant of modulus less than $1$ times $z^t$, and $W_{\mathcal T}>0$ as above. So $k_+\ge1$ and $k_-\ge1$. We construct a Blaschke product that interpolates the data. Put \[ \varphi(\zeta)=\sum_{a\in\mathcal T}\frac{v_a}{1-z_a\zeta},\qquad \psi(\zeta)=\sum_{a\in\mathcal T}\frac{v_af_a}{1-z_a\zeta}. \] The relation $Wv=0$ says that $\varphi(z_b)=f_b\psi(z_b)$ for $b\in\mathcal T$. For $|\zeta|<1$ let $\beta_a=(z_a-\zeta)/(1-z_a\bar\zeta)$. From $(1-z_a\bar\zeta)(1-z_b\zeta)-(z_a-\zeta)(z_b-\bar\zeta)=(1-z_az_b)(1-|\zeta|^2)$ we get \[ \frac{1-|\zeta|^2}{(1-z_a\bar\zeta)(1-z_b\zeta)}=\frac{1-\beta_a\bar\beta_b}{1-z_az_b}. \] Multiplying by $v_av_b(1-f_af_b)$ and summing gives \begin{equation}\label{eq:phipsi} (1-|\zeta|^2)\bigl(|\psi(\zeta)|^2-|\varphi(\zeta)|^2\bigr)=u^*W_{\mathcal T}u,\qquad u_a=v_a\bar\beta_a, \end{equation} because $v^{\mathsf T}W_{\mathcal T}v=0$. So $|\varphi|\le|\psi|$ in the disc. For $|\zeta|=1$ one has \[ \frac1{(1-z_a\bar\zeta)(1-z_b\zeta)}=\frac1{1-z_az_b}\Bigl(\frac1{1-z_a\bar\zeta}+\frac1{1-z_b\zeta}-1\Bigr), \] and summation with $W_{\mathcal T}v=0$ gives $|\varphi|=|\psi|$ on the circle. Since the functions $1/(1-z_a\zeta)$ are independent, $\varphi\ne0$ and so $\psi\ne0$. Hence $h=\varphi/\psi$ is a rational function with real coefficients, with $|h|\le1$ in the disc and $|h|=1$ on the circle. So it is a finite Blaschke product. Both $\varphi$ and $\psi$ are polynomials of degree at most $|\mathcal T|-1$ divided by $\prod_a(1-z_a\zeta)$, so the degree of $h$ is at most $|\mathcal T|-1$. Thus \[ h=\varepsilon z^s\frac{P}{P^*},\qquad \varepsilon=\pm1,\quad P^*(z)=z^{\deg P}P(1/z), \] where $P$ is a real polynomial with all roots in the open disc and $P(0)\ne0$. Put $d=s+\deg P$. Then $d+1\le k_++k_-$. We have $h(z_b)=f_b$ for $b\in\mathcal T$. This is clear if $\psi(z_b)\ne0$. If $\psi(z_b)=0$, then $\varphi(z_b)=0$, and \eqref{eq:phipsi} at $\zeta=z_b$ gives $u^*W_{\mathcal T}u=0$ with $u_b=0$ and $u\ne0$. This contradicts the positive definiteness of $W_{\mathcal T\setminus\{b\}}$. We show that $h$ cannot exist. \emph{The case $t\le s$.} The real polynomial $\varepsilon z^{s-t}P+cP^*$ is nonzero, because $|P|=|P^*|$ on the unit circle and $c<1$. It vanishes at the $k_+$ positive nodes and has degree at most $d-t$. So $k_++t\le d\le k_++k_--1$, that is, $t\frac\pi2k_-$. \emph{The case $s1$ just to the left of $1$. Since $|r|<1$ at the largest node, $r$ takes the value $r(1)=\pm1$ on the right of that node. Rolle's theorem gives one further critical point, and a further root of $R$. The same holds at $-2$. If exactly one of $R(\pm2)$ is negative, this root comes with a sign change of $R$ in $[-2,2]$. If both are negative, there are two such roots. In every case $R$ has two further roots, which contradicts the degree bound. \end{proof} Let $p_1,\ldots,p_m$ be the positive nodes, so $m=k_+$ in the rest of this proof, and let $k=k_-$. We compare the phase of $h$ with the phase of an explicit interpolant of degree $m$. Put $\lambda=c^2$ and $w_a=p_a^t$. Put $\mathsf P_{ab}=(1-\lambda w_aw_b)/(1-p_ap_b)$ and $M=\mathsf P^{-1}$, and write $\langle a,b\rangle=a^{\mathsf T}Mb$. Let $\xi_z(a)=1/(1-zp_a)$, $g_1=1$ and $g_2=-cw$. With $F_i(z)=g_i\xi_z$, $C_j=g_j\xi_{-1}$ and $J=\operatorname{diag}(1,-1)$ define \[ \Theta_{ij}(z)=\delta_{ij}-(1+z)J_j\langle F_i(z),C_j\rangle . \] So $\Theta(-1)=I$. Direct expansion gives the kernel identity \begin{equation}\label{eq:kernelid} \sum_k\Theta_{ik}(z)J_k\Theta_{jk}(u)=\delta_{ij}J_i-(1-zu)\langle F_i(z),F_j(u)\rangle . \end{equation} The vector $F_1(p_a)-f_aF_2(p_a)$ is row $a$ of $\mathsf P$, where now $f_a=-cw_a$. This gives the node relation $\Theta_{1j}(p_a)=f_a\Theta_{2j}(p_a)$. Let $D(z)=\prod_a(1-p_az)$, $b(z)=\prod_a(z-p_a)$ and $\kappa=(-1)^m$. The functions $e_{ij}=D\Theta_{ij}$ are real polynomials of degree at most $m$. Their determinant vanishes at the $p_a$ by the node relation and at the $1/p_a$ because $(e_{ij}(1/p_a))$ has rank one. So it equals $\kappa bD$. Taking $u=1/z$ in \eqref{eq:kernelid} gives \[ e_{11}(z)=\kappa z^me_{22}(1/z),\qquad e_{12}(z)=\kappa z^me_{21}(1/z). \] On the unit circle \eqref{eq:kernelid} with $i=j=2$ and $u=\bar z$ gives $|\Theta_{22}|^2=1+|\Theta_{21}|^2$. Let $\sigma=h(-1)=\pm1$ and $e=\sigma e_{21}+e_{22}$. Then $W=e/D=\Theta_{22}(1+\sigma\Theta_{21}/\Theta_{22})$ does not vanish on the circle, and $W(-1)=1$. \emph{The size of the argument of $\Theta_{22}$.} Let $T_l$ be $\Theta_{22}$ with $\lambda$ replaced by $l$, so $T_0=1$. Let $\cK^{(p)}$ be the matrix with entries $1/(1-p_ap_b)$, let $Z=\operatorname{diag}(w_a)$, and let $M_l=(\cK^{(p)}-lZ\cK^{(p)}Z)^{-1}$. For $0\le l\pi k$ and let $L=\lfloor k/2\rfloor$. If $\mathcal N\ne0$, the intermediate value theorem gives $L$ conjugate pairs of roots on the open arc. So $2L+2m+1\le d+m\le2m+k-1$, which contradicts $2L+2>k$. If $\mathcal N=0$, then $e^{ia}=1$ on the arc, which contradicts the intermediate value theorem at the value $\pi$. Hence $a(\theta_0)\le\pi k$, and \[ 2t\arctan4\le\phi(\theta_0)\le\pi k+2\sum_p\arctan(4\rho_p)+\pi+2\mathcal X . \] This contradicts \eqref{eq:window}, and Theorem~\ref{thm:window} is proved.\qed \subsection{The alternative on near-extremal configurations}\label{ssec:alternative} We now show that \eqref{eq:window} holds with $t=\mathsf e$ and $c=\frac1{100}$ on every configuration for which the energy bound of Section~\ref{sec:energy} does not already suffice. The idea is that such a configuration has Chebyshev moments close to the trial sequence $u$, and the left side of \eqref{eq:window} is then computed by a fixed test function. Let $\phi_H(w)=\frac\pi2-\arctan\bigl(4(1-w)/(1+w)\bigr)$ for $0\le w<1$. Then $\arctan(4\rho_p)=\frac\pi2-\phi_H(p)$. With $\mathsf e=48N-1$, $n=58N$ and $k_-=n-k_+$, the inequality \eqref{eq:window} follows for large $N$ from \begin{equation}\label{eq:phaseneed} \frac1n\sum_{p>0}\phi_H(p)\ge\arctan\tfrac14+\tfrac5{29}\arctan4+\tfrac{17}4\,\frac{c^2}{1-c^2}+\xi,\qquad \xi=\tfrac1{1000}, \end{equation} whose right side is $0.474992\ldots$. Let $f=\frac{F_0}2+\sum_{k\ge1}F_kT_k$ be a polynomial of degree $32$ with dyadic coefficients and denominator $2^{31}$, where $F_0=992223195/2^{30}$. It satisfies $f\le\phi_H$ on $(0,1)$, $f\le0$ on $[-1,0]$, and \begin{equation}\label{eq:testS} \tfrac14F_0+\tfrac14|F_1|y\le\phi_H(y)\qquad(00}\phi_H(p)&\ge\int f\,d(\mu_F+\tfrac12\mu_M)+\tfrac14F_0\beta+\tfrac14|F_1|A_1(\mu_S)\\ &\ge\tfrac12F_0-\tfrac14F_0\alpha+\sum_kF_ku_k(\tau)-\sqrt{\tfrac12\Delta_j\sum_kkF_k^2}, \end{align*} by the Cauchy--Schwarz inequality. The right side is affine in $\alpha$ on each cell. At the end points of every cell it exceeds $0.4752$, which is larger than the right side of \eqref{eq:phaseneed}. The smallest margin is $0.0050$. So \eqref{eq:phaseneed} holds, and Theorem~\ref{thm:window} gives $F(-\frac1{100})\ge0$. All the numerical facts of this subsection are part of the formal proof, which also carries the regularization of the moments that we have suppressed here. We have proved the following. \begin{proposition}\label{prop:alternative} For $N$ large, every configuration with $m\le\alpha_1n$ prescribed columns in the case with doubled interaction, and every $S$, satisfies at least one of two statements. Either the energy bound for $\Xi_S$ gives the conclusion of Proposition~\ref{prop:signed-goal}, or $F(-\frac1{100})\ge0$ and therefore $|\det\Xi_S|\le499^n|\det\Xi^0_S|$. \end{proposition} \subsection{Gram determinants}\label{ssec:gram} It remains to bound $\det\Xi^0_S$. Let $\mathsf D=n-1-g=53N-1$. Let $\xi_1,\ldots,\xi_{k_+}$ be the positive nodes, with $\xi=z$ outside $S$ and $\xi=1/z$ on $S$, so that all these rows are $(\xi^{g-r})_r$. Let $x_1,\ldots,x_{k_-}$ be the negative nodes, and put $u_b=J(x_b)$, where $J(z)=\frac12(z+1/z)$. Let $s=\mathsf D+1-k_-$ and $\omega(z)=\prod_p(z-\xi_p)$. If $k_->\mathsf D+1$, the determinant vanishes, because the negative rows lie in a space of dimension $\mathsf D+1$. So we may assume $s\ge0$. The Laurent polynomials with exponents in $[-\mathsf D,g]$ that vanish at all $\xi_p$ are spanned by the functions $z^{i-\mathsf D}\omega(z)$ with $0\le ig$ the sum in the hypothesis has the single term $d=-a$, and $G_a/A_{-a}\le\frac12$ because $A_a\le A_{-a}$. For $1\le a\le g$ the left side of the hypothesis is $\frac12(A_a+A_{-a})/A_{-a}\le1$. Since $g=5N$ and $10\sigma_j\le\tau_j$, each term of $A_{-a}$ is at most $1$, so $A_{-a}\le16$ for $a\le g$. Also $A_a\ge1$ by the term $j=0$. Hence $\prod_d\theta_d\le16^{5N}$. Let $\tau^*(q)=\max_j(2q\sigma_j-\tau_j)$. Then $G_a\ge\frac14e^{N\tau^*(a/N)}$ for $a\ge1$. Since $\tau^*$ is nondecreasing, \begin{equation}\label{eq:legendre} \sum_{a=0}^{\mathsf D}\log G_a\ge N^2\int_0^{53}\tau^*(q)\,dq-2\mathsf D\log2-N\tau^*(53). \end{equation} The factor $\prod_a\ell_a^2$ equals $2^{\mathsf D(\mathsf D-1)}$. \emph{A tangent bound for Vandermonde determinants.} The integrals $Z_\pm$ are compared with reference measures by the following lemma, which is the negative definiteness of the logarithmic energy on neutral measures in a form exact for finitely many points. Let $d(z,w)^2=|z-w|^2/((1+|z|^2)(1+|w|^2))$ and $\mathsf h(z)=\frac12\log(1+|z|^2)$. For a finite positive measure $q$ on $\C$ write $U^q(z)=\int\log|z-w|\,dq(w)$ and $I(q)=\int U^q\,dq$. Let $H_L=\sum_{j\le L}1/j$. \begin{lemma}\label{lem:tangent} Let $q$ be a finite positive measure on $\C$ with $\int\mathsf h\,dq<\infty$ and $q\{w:d(z,w)^2\le t^4\}\le Ct$ for all $z$ and all $00$ there is $L$ such that for every $m\ge0$, all $\nu_0\ge0$ with $\nu_0q(\C)=m$ and all $z_1,\ldots,z_m\in\C$, \[ |V(z)|^2\le\exp\Bigl(2\nu_0\sum_iU^q(z_i)-\nu_0^2I(q)+2\nu_0m\eta+\tfrac m2H_L\Bigr). \] \end{lemma} \begin{proof} We may assume that the $z_i$ are distinct. Let $\mathsf n(z)$ be the stereographic image of $z$ on the unit sphere. Then $\mathsf C(z,w)=1-d(z,w)^2=\frac12(1+\mathsf n(z)\cdot\mathsf n(w))$ is a sum of four products $b_if_i(z)f_i(w)$ with $b_i\ge0$, and so is every power of $\mathsf C$, with more terms. Let $K_L=-\frac12\sum_{j\le L}\mathsf C^j/j$. Then $K_L=-\sum_\phi b_\phi F_\phi\otimes F_\phi$ is a finite sum with $b_\phi\ge0$. For real $S_\phi=\sum_iF_\phi(z_i)$ and $M_\phi=\int F_\phi\,dq$ the inequality $(S_\phi-\nu_0M_\phi)^2\ge0$ gives, after summation, \begin{equation}\label{eq:gramtangent} \sum_{i,j}K_L(z_i,z_j)\le2\nu_0\sum_i\int K_L(z_i,w)\,dq(w)-\nu_0^2\iint K_L\,dq\,dq . \end{equation} For $z\ne w$ we have $\log d(z,w)=\frac12\log(1-\mathsf C)=-\frac12\sum_{j\ge1}\mathsf C^j/j\le K_L(z,w)$, and $K_L(z,z)=-\frac12H_L$. So $\sum_{i\ne j}\log d(z_i,z_j)$ is at most the left side of \eqref{eq:gramtangent} plus $\frac m2H_L$. Also $\iint K_L\ge\iint\log d$. For the remaining term we show that $\int K_L(z,\cdot)\,dq\le\int\log d(z,\cdot)\,dq+\eta$ for large $L$, uniformly in $z$. Fix $h\in(0,1)$. On $\{d^2\ge h\}$ the difference $K_L-\log d=\frac12\sum_{j>L}\mathsf C^j/j$ is at most $(1-h)^L/(2Lh)$. On $\{d^20$ there is $C$ such that for all $N$, all nodes and all $S$ with $\det\Xi^0_S\ne0$, \[ 2\log|\det\Xi^0_S|\le2\sum_a\mathsf g(a)+\mathsf K_N+4n^2\eta+Cn, \] where $\mathsf g(a)$ is the contribution of node $a$ and \[ \mathsf K_N=2n^2\kappa_++2\cdot53Nn\,c_-+\mathsf D(\mathsf D+1)\log2-N^2\int_0^{53}\tau^*-n^2\bigl(I(q_+)+I(q_-)\bigr). \] \end{proposition} The constant is correct because the $k_-$ points of $Z_+$ give $2nk_-\kappa_+$ and the $s$ points of $Z_-$ give $2nsc_-$. In $\mathsf K_N$ these are written as $2n^2\kappa_+$ and $2\cdot53Nnc_-$, and the differences are absorbed by the terms $-n\kappa_+$ and $-nc_-$ of $\mathsf g$. The term $\mathsf D(\mathsf D+1)\log2$ is the sum of $\mathsf D(\mathsf D-1)\log2$ from $\prod_a\ell_a^2$ and $2\mathsf D\log2$ from \eqref{eq:legendre}. \subsection{The final inequality}\label{ssec:budget} We now bound the other factors of the integrand. Let $n'=n-m$, with free points $x_i$ and $s_j$ and prescribed points $y_j$, and let $\mathcal S$ be the product of the factors of the integrand other than the row determinant and the factors $\mathcal B_\zeta$ of Lemma~\ref{lem:binomvand}. Put $\mathsf t(x)=2x/(1+x^2)$ and $t_i=\mathsf t(x_i)$. In the normalization of \S\ref{ssec:rowdet} a free point carries the weight $|t|^{75N-1}(1-t)^{3N}$. The prescribed columns have lost the common factor $|t|^{22N}(1-t)^{3N}$ of the rows, so a prescribed point carries $|\mathsf t(y)|^{\mathsf D}$. The Cauchy determinant of Lemma~\ref{lem:andreief} is $\pm V(t)V(s)/\prod_{i,j}(1-t_is_j)$, and the generalized Vandermonde determinant gives $\prod_js_j^{9N}(1-s_j)^{3N}|V(s)|$. Hence \begin{multline*} \mathcal S=\prod_j|\mathsf t(y_j)|^{\mathsf D}\prod_i|t_i|^{75N-1}(1-t_i)^{3N}\\ \cdot\prod_js_j^{9N}(1-s_j)^{3N}\,\frac{|V(t)|\,V(s)^2}{\prod_{i,j}(1-t_is_j)}. \end{multline*} Now $1-t=(1-x)^2/(1+x^2)$, $1-ts=(1-2xs+x^2)/(1+x^2)$ and \[ t_i-t_{i'}=\frac{2(x_i-x_{i'})(1-x_ix_{i'})}{(1+x_i^2)(1+x_{i'}^2)}. \] Each $x_i$ occurs in $n'-1$ differences and in $n'$ kernels, so the exponent of $1+x_i^2$ is $-(75N-1)-3N-(n'-1)+n'=-(78N-2)$. Taking logarithms gives \begin{multline}\label{eq:scalar} \log\mathcal S=\Bigl(n'(75N-1)+\binom{n'}2\Bigr)\log2+\sum_j\mathsf w(y_j)+\sum_i\mathsf W_N(x_i)\\ +\sum_j\bigl(9N\log s_j+3N\log(1-s_j)\bigr)+n^2\mathsf Q(x,s), \end{multline} where $\mathsf w(y)=\mathsf D(\log2+\log|y|-\log(1+y^2))$, $\mathsf W_N(x)=(75N-1)\log|x|+6N\log(1-x)-(78N-2)\log(1+x^2)$ and \begin{multline*} n^2\mathsf Q(x,s)=\sum_{i0$ there is $\nu_1$ such that for all $n\ge\nu_1$, all $n'\le n$ and all configurations of distinct points, \begin{multline*} \mathsf Q(x,s)\le\tfrac32b^2\log2+\|a\|^2+\tfrac12\|v\|^2-\tfrac32\bigl(\tfrac{n'}n-b\bigr)^2\log2\\ +\frac1n\sum_i\mathsf R_*(x_i)+\frac1n\sum_j\mathsf C_*(s_j)\\ -\frac1{2n^2}\sum_i\log(1-x_i)-\frac1{n^2}\sum_j\log(1-s_j)+\varepsilon, \end{multline*} where $\mathsf R_*(x)=-b\log2-2\mathsf T(a,x)-\mathsf S(v,x)$ and $\mathsf C_*(s)=-2b\log2+2\mathsf T(v,s)$. \end{lemma} \begin{proof} Use the expansions $-\log|x-x'|=\log2+2\sum_kT_k(x)T_k(x')/k$, $-\log(1-xx')=\sum_kx^kx'^k/k$ and $-\log(1-2xs+x^2)=2\sum_kx^kT_k(s)/k$. Let $A_k=\frac1n\sum_iT_k(x_i)$, $B_k=\frac1n\sum_ix_i^k$ and $C_k=\frac1n\sum_jT_k(s_j)$, and let $f=n'/n$. If the diagonal terms are ignored, these expansions give \[ \mathsf Q=-\tfrac32f^2\log2-\|A\|^2-\tfrac12\|B-2C\|^2 . \] The diagonal terms diverge. OpenAI's regularization \cite[\S5]{OpenAI2026} replaces the kernels by Abel-damped series with a parameter $e$. It gives the identity above for the damped moments, up to an error at most $-\frac1{2n^2}\sum_i\log(1-x_i)-\frac1{n^2}\sum_j\log(1-s_j)+O_e(1/n)+14e$. For the trial sequences, $-\|A\|^2\le\|a\|^2-2\sum_ka_kA_k/k$ and $-\frac12\|B-2C\|^2\le\frac12\|v\|^2-\sum_kv_k(B_k-2C_k)/k$. Removing the damping from the linear terms costs $O(e)$ times the $\ell^1$ norms of the trial sequences, so we first replace $a$ and $v$ by long prefixes, with an error at most $\varepsilon/4$ uniformly in the points. Finally $-\frac32f^2\log2=-\frac32(f-b)^2\log2+\frac32b^2\log2-3bf\log2$, and the last term is split between $\mathsf R_*$ and $\mathsf C_*$. Choosing the prefix, then $e$, then $\nu_1$ proves the lemma. The order of these choices makes the bound uniform in the configuration. \end{proof} We apply the lemma with $b=1$. Let $X$ be the inverse of $x\mapsto x+1/x$ with values in $[-1,1]\setminus\{0\}$. Let $R$ and $C$ be the row and column references of Appendix~\ref{app:trials}, and let $L(\mu)=\int\log|X|\,d\mu$. We take \[ a_k=\int T_k(X)\,dR,\qquad v_k=\int X^k\,dR-2\int T_k\,dC, \] and put $\mathsf I_0=\frac32\log2+\|a\|^2+\frac12\|v\|^2$. The expansions in the proof of Lemma~\ref{lem:freetangent} give \begin{gather*} \mathsf R_*(x)=\int\bigl(\log|x-X|+\log(1-xX)\bigr)\,dR-\int\log(1-2xs+x^2)\,dC(s),\\ \mathsf C_*(s)=2U^C(s)-\int\log(1-2Xs+X^2)\,dR. \end{gather*} For nonzero $x$ and $w$ and for $00$, so $-\log(1-xX)=\sum_kx^kX^k/k$ has positive terms. Those with $k\le900$ integrate to $\mathsf P(x)$ and the others to at most $\mathsf E$. Hence \begin{equation}\label{eq:qplus} U^{q_+}(x)\le L(G_+)+U^{G_+}(Y)+m_+\log x+\mathsf P(x)+\mathsf E+A_+(x). \end{equation} Let now $-10$. Add $\mathsf W_N$ to this and use \eqref{eq:qplus} for $00$ outside $S$, \eqref{eq:qplus} and $53N=\mathsf D+1$ give \[ \mathsf w(y)+\mathsf g(y)\le \mathsf D\log2+n\mathsf H_+(y)+\log(1+y^2). \] For $y<0$, \eqref{eq:gneg} gives \[ \mathsf w(y)+\mathsf g(y)=\mathsf D\log2+n\mathsf H_-(y)-\log|y|+\log(1+y^2). \] For $y\in S$ the base point is $\xi=1/y$, so $\mathsf g=\mathsf D\log y+n(U^{q_+}(1/y)-\kappa_+)$. We have $\log|1/y-w|=-\log y+\log|1-yw|$, the measure $q_+$ has mass $1$, and $\int\log(1-yX)\,dG_+\le-\mathsf P(y)$ because all the moments are nonnegative. Hence $U^{q_+}(1/y)\le-\log y-\mathsf P(y)+\int\log|1-yw|\,d\nu_+(w)$. With $2\mathsf D=106N-2$ this gives \[ \mathsf w(y)+\mathsf g\le \mathsf D\log2+n\mathsf H_S(y)-2\log y+\log(1+y^2). \] The factor $e^{-N\hat\delta(\mathsf t(y))}$ of Proposition~\ref{prop:signed-goal} adds $\hat\delta(\mathsf t(y))/58$ to these fields. Let $\mathsf M$ be an upper bound for $\mathsf H_\pm+\hat\delta\circ\mathsf t/58$ on $E_W$ and for $\mathsf H_S+\hat\delta\circ\mathsf t/58$ on $E_W\cap(0,1)$. Since $\log(1+y^2)\le\log2$ and $|y|$ is bounded below on $E_x$, every prescribed row, with its factor $e^{-N\hat\delta}$, contributes at most $\mathsf D\log2+n\mathsf M+O(1)$. Now add \eqref{eq:scalar}, Lemma~\ref{lem:freetangent} multiplied by $n^2$, and half of Proposition~\ref{prop:gas}. With $n'=(1-\alpha)n$ and $\eta,\varepsilon\to0$, this gives \[ \log\bigl(\mathcal S\,|\det\Xi^0_S|\bigr)\le n^2\bigl(\mathsf B(\alpha)+o(1)\bigr), \] uniformly in the configuration, with \begin{multline}\label{eq:budget} \mathsf B(\alpha)=\mathsf I-\tfrac32\alpha^2\log2+(1-\alpha)(\mathsf R+\mathsf C_0)+\alpha\mathsf M\\ +\log2\Bigl(\tfrac{75}{58}(1-\alpha)+\tfrac12(1-\alpha)^2+\tfrac{53}{58}\alpha\Bigr). \end{multline} Here $\mathsf I=\mathsf I_0(\tau)+\mathsf I_{\mathrm{gas}}$, and $\frac12\mathsf K_N/n^2$ tends to \[ \mathsf I_{\mathrm{gas}}=\kappa_++\tfrac{53}{58}(\kappa_--c_M)-\tfrac12\bigl(\tfrac{53}{58}\bigr)^2\log2-\tfrac12\bigl(I(q_+)+I(q_-)\bigr)-\frac1{2\cdot58^2}\int_0^{53}\tau^*, \] with $\kappa_-$ and $c_M$ as above. The three terms in $\log2$ of \eqref{eq:budget} come from the exponent $n'(75N-1)+\binom{n'}2$ and from the term $\mathsf D\log2$ of each prescribed row. Appendix~\ref{app:trials} gives the reference measures and the bounds \begin{equation}\label{eq:caps} \begin{gathered} \mathsf I_0+\mathsf I_{\mathrm{gas}}\le2.366997,\qquad\mathsf M\le-0.976,\\ c^{\mathsf R}_0=-2.579740254,\quad c^{\mathsf R}_1=-2.574009165,\\ c^{\mathsf C}_0=-2.946694099,\quad c^{\mathsf C}_1=-2.888907917,\quad c^{\mathsf C}_2=-2.85421372, \end{gathered} \end{equation} with $\kappa_-=5463909/10^8$ and $\kappa_+=14762657/10^8$, and the moved values \begin{gather*} L_0\le-7.5421576\cdot10^{-5},\quad L_1\le-0.04943292757,\\ Q_{00}\le1.6623\cdot10^{-7},\quad Q_{01}\le9.5232\cdot10^{-7},\quad Q_{11}\le1.002863\cdot10^{-3}. \end{gather*} Let $V_S(\alpha)=\mathsf B(\alpha)+U_0-\frac12\log2-c_2\alpha$, with $\mathsf I$, $\mathsf R$ and $\mathsf C_0$ the moved values above and $\log2$ replaced by an upper bound. On each of the two pieces $[0,\alpha_e]$ and $[\alpha_e,\alpha_1]$ it is a polynomial in $\alpha$ with rational coefficients. We have $V_S(0)=-1.81\ldots\cdot10^{-5}$, and the cover of \S\ref{ssec:cases} proves $V_S(\alpha)\le\sigma\alpha$ on both pieces. The inequality is tight near $\alpha=0.0033$, where $V_S(\alpha)/\alpha$ reaches $0.33867$. \begin{proof}[Proof of Proposition~\ref{prop:signed-goal}] Fix a configuration with $m\le\alpha_1n$. By \eqref{eq:expandS} it suffices to bound each of the at most $2^m$ terms $\mathcal S\,|\det\Xi_S|$, at the cost of the factor $5^m$. By Proposition~\ref{prop:alternative} each term either satisfies the bound directly, or is at most $499^n\mathcal S\,|\det\Xi^0_S|$. In the second case the final inequality gives $\exp\bigl(n^2(-U_0+\frac12\log2+c\alpha+o(1))-N\sum_j\hat\delta(t_{a_j})\bigr)$, since $499^n=e^{o(n^2)}$ and $c=c_2+\sigma$. The factor $\prod_l(1+\zeta s_l)^m$ of Lemma~\ref{lem:binomvand} is contained in the column field, and the factors $\mathcal B_\zeta$ give $\prod_j\mathcal B_\zeta(k_j)$. \end{proof} \section{Proof of Theorem~\ref{thm:main}}\label{sec:proof} OpenAI's energy method gives the decay rate $D_0=n_0^2(U-\frac12\log2)=\DecayOA\ldots$ for our parameters. Section~\ref{sec:signed} raises it to $D=\Decay\ldots$ With $\Lambda=\LamDen\ldots$ the polynomial $Q_N=(d_0N)!\,\cD_N\Delta_N$ has integer coefficients, degree at most $53N$, and \[ |Q_N(x)|\le e^{-\eta N^2}\qquad(|x-G|\le e^{-\rho N}),\qquad \eta=\EtaTot . \] Two further properties are needed to choose the prime in the criterion: nonvanishing of $Q_p(a/b)$ when $p^2\nmid b$ (Proposition~\ref{prop:nonvanishing}), and divisibility of the $i$-th coefficient of $Q_N$ by $p'^{2i}$ for every prime $N1$. Finally $d+4d\delta\lambda=\PayL\ldots<\eta\lambda=\PayR\ldots$ So Proposition~\ref{prop:criterion} gives $|G-a/b|>b^{-\rho\lambda}=b^{-270}$ for all integers $a$ and all sufficiently large $b$. This proves Theorem~\ref{thm:main}. \section{Remarks on the method}\label{sec:remarks} \subsection{The three ingredients} For fixed parameters the method gives the exponent $d\rho/\eta$, and its three ingredients come from different places. The degree $d=n-g$ comes from the construction. Of $\rho=\Rho$, the prescribed columns contribute $n_0c\approx228.3$ through the energy, and the weight and the Vandermonde factor of one column contribute $\psi\approx50.1$. Of $n_0c$, the part $n_0\sigma\approx19.6$ comes from the bound of Section~\ref{sec:signed} for few prescribed columns. The number $\eta$ is a small difference of two large numbers, $D\approx6.4\cdot10^3$ and $\Lambda\approx6.4\cdot10^3$. The corank bounds of Section~\ref{sec:content} contribute about $22$ to it, and the bound of Section~\ref{sec:signed} about $25$. Without them $\eta$ would be about $8$. Since the exponent is inversely proportional to $\eta$, a further gain of $0.2$ in $\eta$ would lower it by about $1$. \subsection{The row determinant} \begin{remark}\label{rem:L} The energy method bounds $|\det\mathsf L|$ by $e^{O(n\log n)}$, which is invisible at the scale $e^{N^2}$. At the extremal configuration, numerical evaluation for small $N$ suggests that $\log|\det\mathsf L|$ is about $-28N^2$. So our bound captures most of its true size. The comparison of Lemma~\ref{lem:compare} loses only $e^{O(N)}$, and so do the Gram inequalities of Proposition~\ref{prop:gram}. The loss is in the restriction to Gram measures on finitely many circles, and in the requirement that the bound hold for the maximum over all configurations. These numerical values are not part of the formal proof. \end{remark} \subsection{Other parameters} All the results of Section~\ref{sec:content} are lower bounds for the corank. The closed formulas, the counting inequality, Theorem~\ref{thm:everyU} and the families derived from them are proved for all parameters. The rank data of Appendix~\ref{app:windows} and the reference measures belong to $(58,9,3,5,3)$, and Sections~\ref{sec:real}--\ref{sec:signed} are written and formalized for these parameters. Theorem~\ref{thm:window} does not depend on the parameters. A different choice of parameters changes $d$, $\rho$ and $\eta$ together, and we make no claim that ours is optimal. Except for the pair (F4) and the intervals of Theorem~\ref{thm:everyU}, the function $\kappa$ uses no relation between different families. So at a prime where intervals overlap the exact corank may be larger than each bound. An upper bound for the corank is not needed. More independence between families would increase $\eta$ without changing anything else. \section{Formalization}\label{sec:formal} Theorem~\ref{thm:main} and the results it depends on are formalized in Lean~4 \cite{Lean4} with Mathlib \cite{Mathlib}, on top of OpenAI's formalization of the irrationality of $G$ \cite{OpenAI2026}. The development is available at \begin{center} \FormalizationLink. \end{center} Theorem~\ref{thm:main} is the theorem \begin{center} {\small\texttt{CatalanIrrationalityMeasure.catalan\_irrationalityMeasure\_270}} \end{center} in \lean{Solution.lean}. Its statement, in \lean{Challenge.lean}, uses only Mathlib. Catalan's constant is written as the series $\sum_j(-1)^j/(2j+1)^2$, and the conclusion is $|G-a/b|>b^{-270}$ for all integers $a$ and all sufficiently large $b$. The modules of the proof contain no \lean{sorry} and no \lean{native\_decide}. The command \lean{\#print axioms} shows that only the three standard axioms \lean{propext}, \lean{Classical.choice} and \lean{Quot.sound} are used. The proof is kernel-checked: Lean's kernel accepts every declaration. It has not been run through the independent checker \lean{lake comparator} \cite{Comparator}. The prime number theorem is taken from the PNT+ project \cite{PNTPlus}. The parts of OpenAI's formalization that are used were converted to Lean's module system and ported to Lean 4.35.0-rc2. The repository's \texttt{README} describes the changes and the layout of the modules. Two lemmas of Section~\ref{sec:real} on the weights of nodal polynomials come from Davis's formalization \cite{Davis2026}. We use two of its Lean files, taken from commit \texttt{3c45153} under the Apache License~2.0. Their declarations and proofs are unchanged. We adapted the module paths and imports, and added comments and axiom checks. The repository's \texttt{NOTICE} file and a provenance record list these files with their source hashes. The proof combines conventional formal proofs with numerical certificates. These are the inverses of the base matrices modulo $1009$, the intervals and the chord table of Section~\ref{sec:real}, the trial sequences, the cell covers and the bounds of Appendix~\ref{app:field}, the reference measures and the bounds of Appendix~\ref{app:trials}, and the rank data of Appendix~\ref{app:windows}. The certificates are exact rational and integer computations checked by the Lean kernel without compiled code. The interval arithmetic used to generate them \cite{Arb} is not trusted. \section*{Disclosure of AI use} The mathematics, the Lean formalization, the certificate generators and the text of this paper were produced by AI systems at the request of the author. The author posed the problems, directed the work and is responsible for the paper. The systems were Claude Opus~5.5, Claude Sonnet~5.5 and Claude Haiku~5.5 (Anthropic), running in Claude Code, and GPT-6.1 Sol (OpenAI), running in Codex. The underlying irrationality proof for Catalan's constant and its Lean formalization are due to OpenAI \cite{OpenAI2026}. \subsection*{Acknowledgements} We thank the authors of Mathlib and of the PNT+ project, OpenAI for making its formalization available, and Cinna Davis for making the formalization of \cite{Davis2026} available. \appendix \section{Counting on every subspace}\label{app:windows} On $37$ of the $113$ intervals of Theorem~\ref{thm:profile-D} the corank bound comes from the counting inequality applied to all constant subspaces at once. These intervals lie in $(3N,16.2N)$. No subspace is constructed. The proof bounds the maximum over subspaces of the right side of Theorem~\ref{thm:everyU} by forced intersections of row spaces, and the only finite data are upper bounds for ranks. Let $V$ be a finite-dimensional vector space over $\F_p$. Let $c$ be a nonnegative integer. For $j$ in a finite set $\mathcal J$ let $B_j\colon V\to W_j$ be a linear map and $w_j$ a nonnegative integer, and let $\lambda_{j,k}$ ($kq$, we have $a=0$ in Theorem~\ref{thm:count}. \emph{Lower bounds for the maximum.} For a subspace $S$ of the dual $V^*$ let $U=S^\perp$. Let $R_j\subset V^*$ be the row space of $B_j$, and let $r_j\ge\dim R_j$ be real numbers, in practice upper bounds for the ranks. Since $\rank(B_j|_U)=\dim R_j-\dim(R_j\cap S)$, the right side of Theorem~\ref{thm:everyU} is \[ E(S)=c(\dim V-\dim S)-\sum_jw_j\bigl(\dim R_j-\dim(R_j\cap S)\bigr). \] Let $e_0=c\dim V-\sum_jw_jr_j$ and $\delta_j=r_j-\dim R_j\ge0$. Then \begin{equation}\label{eq:eulerid} E(S)=e_0+\sum_jw_j\delta_j+\sum_jw_j\dim(R_j\cap S)-c\dim S . \end{equation} A map may also be omitted from the sum if its weight is subtracted from $c$, because $\dim R_j-\dim(R_j\cap S)\le\dim V-\dim S$. The following three lemmas give a subspace with $E(S)>e_0$, from relations among the row spaces. All sets of indices are subsets of $\mathcal J$, and all $w_j$ are nonnegative. \begin{lemma}[pair]\label{lem:pairU} Let $a\ne b$ lie in a set $I$ such that $R_a\le R_j$ or $R_b\le R_j$ for all $j\in I$. Let $s\ge\dim(R_a+R_b)$ and $0\le h\le\min\bigl(w_a,w_b,\sum_{j\in I}w_j-c\bigr)$. Then some $S$ has $E(S)\ge e_0+\max(0,r_a+r_b-s)\,h$. \end{lemma} \begin{proof} Take $S=R_a\cap R_b$ and let $k=\dim S$. Then $S\le R_j$ for $j\in I$, so \eqref{eq:eulerid} gives \[ E(S)-e_0\ge w_a\delta_a+w_b\delta_b+\Bigl(\sum_{j\in I}w_j-c\Bigr)k\ge h(\delta_a+\delta_b+k). \] Since $\dim(R_a+R_b)+k=\dim R_a+\dim R_b$, we have $r_a+r_b-s\le\delta_a+\delta_b+k$. Also $\delta_a+\delta_b+k\ge0$. \end{proof} \begin{lemma}[several points]\label{lem:pointsU} Let $c\ge0$. Let $I_1,\ldots,I_l$ be disjoint sets with $a_i\in I_i$, such that $R_{a_i}\le R_j$ for $j\in I_i$. Put $W_i=\sum_{j\in I_i}w_j$ and suppose that $W_i\le c$ for each $i$. Let $s\ge\dim\sum_iR_{a_i}$. Let $h\ge0$ with $h\le w_{a_i}$ for each $i$ and $h\le\sum_iW_i-(l-1)c$. Then some $S$ has $E(S)\ge e_0+\max(0,\sum_ir_{a_i}-s)\,h$. \end{lemma} \begin{proof} Let $\mathsf T$ be the kernel of the sum map $\bigoplus_iR_{a_i}\to V^*$, and let $k=\dim\mathsf T$. Let $K_i\le R_{a_i}$ be the projection of $\mathsf T$ to the summand $i$, let $\kappa_i=\dim K_i$, and let $S=\sum_iK_i$. Then $\kappa_i\le k$. The rank formula for the sum map gives $\sum_i\dim R_{a_i}\le s+k$. The kernel of the sum map on $\bigoplus_iK_i$ is again $\mathsf T$, so $\dim S=\sum_i\kappa_i-k$. For $j\in I_i$ we have $K_i\le R_j\cap S$, so $\dim(R_j\cap S)\ge\kappa_i$. Since the $I_i$ are disjoint, the $a_i$ are distinct, and \eqref{eq:eulerid} gives \[ E(S)-e_0\ge\sum_iw_{a_i}\delta_{a_i}+\sum_iW_i\kappa_i-c\sum_i\kappa_i+ck\ge h\sum_i\delta_{a_i}+\sum_i(W_i-c)\kappa_i+ck. \] Since $W_i-c\le0$ and $\kappa_i\le k$, the last two terms are at least $k(\sum_iW_i-lc+c)\ge hk$. So $E(S)-e_0\ge h(\sum_i\delta_{a_i}+k)$, and $\sum_ir_{a_i}-s\le\sum_i\delta_{a_i}+k$. \end{proof} \begin{lemma}[kernel]\label{lem:kernelU} Let $R\subset V^*$ be a subspace with $\dim R\le s$, and let $Z$ be a set with $R_j\le R$ for $j\in Z$. For every integer $d\ge0$ with $d+s\le\dim V$ some $S$ has $E(S)\ge cd-\sum_{j\notin Z}w_j\min(r_j,d)$. \end{lemma} \begin{proof} The space $R^\perp\subset V$ has dimension at least $d$. Choose $U\le R^\perp$ of dimension $d$ and let $S=U^\perp$. Then $B_j|_U=0$ for $j\in Z$, and $\rank(B_j|_U)\le\min(r_j,d)$ for the other $j$. \end{proof} \emph{The rank data.} The space $V$ is a space of profiles of the coefficient vectors of a block, with coordinates $T_k$. The conditions at the moment columns are linear combinations of a small number of basic rows \[ A[k]\mapsto T_k+T_{k+1},\qquad C[k]\mapsto T_{k+1}-T_k, \] together with the row of the centre coordinate. So the moment values themselves are not needed. Every row factors through the map $G$ whose components are these basic rows. This gives $r_j\le\min(\rank F_j+\#\text{basic rows},\ \rank(G,F_j))$, where $F_j$ is the fixed part of the map. Each rank bound is certified by an integer left inverse, checked by the Lean kernel. Integer left inverses are valid at every prime, so no prime enters this step. The $37$ intervals are treated as follows: $13$ by Lemma~\ref{lem:kernelU}, $12$ by Lemma~\ref{lem:pointsU}, $11$ by Lemma~\ref{lem:pairU}, and one by $E(V)$ itself. On each interval the bound is an affine function $\nu_N(p)\ge c_0N+c_1p-2$ with rational $c_0$ and $c_1$. The rounding of the integer orders is absorbed by the shortening $\tau$ of the interval and the lowering of the bound by $10^{-6}$ in Section~\ref{ssec:kappa}. \section{The bounds for the energy method}\label{app:field} This appendix gives the trial sequences of Section~\ref{sec:energy}, the certified bounds that they satisfy, and the bounds for the weights of Section~\ref{sec:real}. \emph{Trial sequences.} For the constrained case we use a trial sequence with $40$ Chebyshev coefficients and a finite number of geometric terms, with the multiplier $\lambda_1=1.68493357$. For the case with doubled interaction we use trial sequences of the form \eqref{eq:trial}, with $\Kcoef$ Chebyshev coefficients, which are dyadic rationals, and $\Ngeo$ geometric terms. Of these, $\Ncomplex$ have complex $z_j$ and complex $r_j$. We place the complex charges at the $41$ equally spaced points $x_0=-0.16+0.008j$ $(0\le j\le40)$, all with $\epsilon_0=10^{-2}$. The real terms sit at points $\pm z$ with $z$ increasing from $0.8$ to $0.999997$. The numbers $z_j$ are rationals with at most twelve decimals. The norms of the geometric terms have closed forms: two terms $(z_j,r_j)$ and $(z_l,r_l)$ contribute $-\frac12\operatorname{Re}\bigl(r_jr_l\log(1-z_jz_l)+r_j\bar r_l\log(1-z_j\bar z_l)\bigr)$ to $\|u\|^2$. \emph{The suprema of the fields.} We must bound two functions of one variable from above on their domains: the field $X$ on $[-1,1]$ and the field $Y$ on $[0,1]$. We locate the real critical points. The derivative is a rational function with a known denominator, and an exact identity gives its numerator. The Descartes rule, after a M\"obius transformation of the interval, bounds the number of real roots of the numerator. Sign evaluations at $81$-bit rational points then isolate this number of roots. One global Lipschitz bound on each isolating interval and the values at the end points give the supremum. \emph{The bounds.} For the base trial sequences this gives \[ \sup X\le\SupX,\quad \sup Y\le\SupY . \] So $M_X+M_Y=\FieldCap$, and the norms satisfy $2\|u\|^2+\frac12\|v\|^2\le0.7736206$. The first coefficient satisfies $u_1\ge9/50$, which is used near $x=0$. For the constrained case, with the weights $(\alpha_A,\beta_h,\gamma_H-\frac12,\lambda_1,2g_0/n_0)=(11/29,3/29,49/58,\lambda_1,5/29)$, the bounds $\sup X_1\le-1.15854$ and $\sup Y\le-1.32981$ are obtained by isolating the zeros of the numerators of the derivatives, which are polynomials with integer coefficients. The norm bound is $n_1=0.67761$. \emph{The increments.} The increment of the case with doubled interaction has $40$ Chebyshev coefficients, and that of the constrained case has $24$. Both are dyadic. Their fields are bounded on the whole domains by the same methods. For the case with doubled interaction the norm bound at $\tau$ is $0.7736206-0.117647903\,\tau+0.005215061\,\tau^2$, the field $X$ gains at most $0.01484272\,\tau$, and the column field $Y$ with $\gamma g$ is bounded by a piecewise linear function of $\gamma$ with knots at $\alpha_1,\frac15,\frac3{10},\frac25,\frac35,\frac45,1$. The fields $F_1$ and $F_2$ contain the node term $\hat\delta/58$. They are bounded by covers by cells, with exact rational and integer arithmetic on each cell and the maximum of $\hat\delta$ on each cell. For $F_1$ on $|x|\ge401/2500$ the cover has $150$ cells. It uses a Taylor expansion of the polynomial part and an expansion to order $5$ of the logarithmic terms. On the positive part of $E_W$ below $401/2500$ six further cells suffice. For $F_2$ the covers of $[98749/10^6,\,9995/10^4]$ have $17$ and $14$ cells at the two ends of the move. This gives \[ \sup F_1\le0.007,\qquad \sup F_2\le0.007-\log2 \] at $\tau=0$, and $-0.0268$ and $0.0233-\log2$ at $\tau=1$. For the constrained case the norm bound at $\tau$ is $0.67761-0.082834152\,\tau+0.004096781\,\tau^2$, and the bound $\mathcal M_1$ of Section~\ref{sec:energy} is the chord of $0.35072$ and $0.35295$. The two numbers are maxima over the $865$ segments of the chord table at the two end values of $\lambda_1$. With $a_0=41/58$, $a_1=76/29$ and $U_0=\Uzero$ these bounds give the inequalities of \S\ref{ssec:cases}. Each check $V(\alpha)\le\sigma\alpha$ is a polynomial inequality on a closed interval and is proved by a cover in exact rational arithmetic. \emph{The bound for the weights.} Inequalities \eqref{eq:circlecap} and \eqref{eq:segcap} concern the explicit function $L$ of \eqref{eq:Ldef}. On the circle $|z|=\frac34$ write $z=\frac34(s+i\sqrt{1-s^2})$ with $s=\cos\theta$. Then $|z-t|^2=\frac9{16}+t^2-\frac32ts$ is affine in $s$ for every real $t$. So $\log|1-z|$ is a concave function of $s$. Each $G_c(z)$ is $\operatorname{arccosh}$ of $(|z-a_c|+|z-b_c|)/w_c$, a sum of square roots of affine functions of $s$, and therefore concave in $s$ as well. Hence the left side of \eqref{eq:circlecap} is a concave function of $s\in[-1,1]$. Its maximum is $3.25053\ldots$ near $s=-0.1741$. A concave function is bounded by its secants outside the points and by the chord between them, so three points at distance $\frac1{2000}$ around the maximum suffice. At these points the value is enclosed to within $10^{-17}$ by exact rational arithmetic with integer square roots, and the bound $D_\Gamma=\Dcirc$ holds with a margin $6\cdot10^{-5}$. On the real segments we use, for real $x$, the closed form $e^{G_c(x)}w_c/4\le\frac12\max(w_c,|2x-a_c-b_c|)$ together with $6\log|1-x|\le6\log\frac14$ for $x\ge\frac34$ and $6\log|1-x|\le6\log2$ for $x\le-\frac34$. This gives \eqref{eq:segcap} with values below $-9$, far below $D_\Gamma$. The knot values $v_i$ of the chord table lie between $-1.97$ and $17.99$. \section{The reference measures of Section~\ref{sec:signed}}\label{app:trials} The bound of \S\ref{ssec:budget} uses reference measures $q_+$ and $q_-$, a row reference $R$ and a column reference $C$. This appendix describes them and the certified bounds \eqref{eq:caps}. \emph{The measures.} On the real line of the variable $Y=x+1/x$ with $|Y|\ge2$ we take measures $G_+$ and $G_-$ that are piecewise uniform on finitely many cells in $Y$, with masses $m_+$ and $m_-$ and total mass $1$. They carry the positive and the negative free nodes, and $R=G_++G_-$ is the row reference. The column reference $C$ has mass $1$. The measures are completed by auxiliary measures on the $16$ inner circles. Each auxiliary measure has a cubed Poisson density: its Fourier coefficients are $-q^k(1+Ak+Bk^2)/k$ with $q=e^{-\varepsilon/2}$ and $\varepsilon=27/250$, and the constants $A$ and $B$ are rational functions of $r=q^2$. The auxiliary measures $\nu_+$ and $\nu_-$ sit at the phases $\pi j/93$, with $616$ and $935$ entries respectively, and have total masses $1-m_+$ and $\frac{53}{58}-m_-$. Let $X$ be the inverse of $x\mapsto x+1/x$ with $|X|\le1$. Then $q_+=X_\#G_++\nu_+$ is the measure in the plane of $z$. In the plane of $u$ we take $q_-=J_\#\tilde q_-$ with $\tilde q_-=X_\#G_-+\nu_-$. \emph{The fields.} Let $L(\mu)=\int\log|X|\,d\mu$. For $\sigma=\pm$ the free row field is \begin{multline*} F_\sigma(x)=L(R)+L(G_\sigma)-\kappa_\sigma+U^{G_\sigma+R-2_\#C}(x+1/x)+\bigl(m_\sigma+\tfrac{11}{29}\bigr)\log|x|\\ -\tfrac{39}{29}\log(1+x^2)+\tfrac3{29}\log(1-x)+[\sigma=+]\bigl(\mathsf P(x)+\mathsf E\bigr)+A_\sigma(x), \end{multline*} where $A_+=U^{\nu_+}$ and $A_-(x)=\int(\log|x-w|+\log|1-xw|)\,d\nu_-(w)$. The polynomial $\mathsf P(x)=\sum_{k\le900}m_kx^k/k$ with $m_k=\int X^k\,dG_+$ is the truncated reflection potential of the positive measure, and $\mathsf E\ge0$ bounds its tail. The column field is \[ \mathsf C(s)=-L(R)+2U^C(s)-U^R(2s)+\tfrac9{58}\log s+\tfrac3{58}\log(1-s). \] A prescribed row outside $S$ has the field $\mathsf H_\sigma$ obtained from $F_\sigma$ by three changes. The references $R$ and $C$ are removed, the coefficients $\frac{11}{29}$ and $\frac3{29}$ are replaced by $0$, and $\frac{39}{29}$ is replaced by $\frac{53}{58}$. Its domain is $x\le-19501/10^5$ and $x\ge98749/10^6$, which contains $E_x$. A prescribed row in $S$ has the field \[ \mathsf H_S(y)=\tfrac{48}{58}\log y-\tfrac{53}{58}\log(1+y^2)-\kappa_+-\mathsf P(y)+\int\log|1-yw|\,d\nu_+(w) \] for $0