\begin{filecontents*}[overwrite]{leannote.sty} \NeedsTeXFormat{LaTeX2e} \ProvidesPackage{leannote}[2026/11/18 Lean margin notes] \RequirePackage{xcolor,url,needspace,zref-savepos,zref-abspage,etoolbox} \makeatletter \newcommand*\leannotefont{\fontfamily{zi4}\fontseries{m}\fontshape{n}\selectfont} \newcommand*\leannotecolor{\color[gray]{0.33}} \newcommand*\leannotestyle{% \everypar{}\normalfont\scriptsize\upshape \setlength\baselineskip{9.2pt}% \parindent\z@\leftskip\z@\rightskip\z@\@plus 2.5em\relax \parfillskip\z@\@plus 1fil\relax\spaceskip\fontdimen2\font\xspaceskip\z@ \hyphenpenalty 4000\exhyphenpenalty 4000\pretolerance\m@ne\tolerance 9000\relax \emergencystretch 1em\hbadness 10000\relax \let\nolinkurl\leandecl\let\url\leandecl} \DeclareUrlCommand\leandecl{% \def\UrlFont{\leannotefont\color[gray]{0.18}}% \def\UrlBreaks{\do\.\do\_}\def\UrlBigBreaks{}\def\UrlNoBreaks{}% \def\UrlSpecials{}\Urlmuskip=0mu plus 1mu\relax} \@namedef{PFln@lead@same}{Same route.} \@namedef{PFln@lead@formal}{Formalized.} \@namedef{PFln@lead@route}{Different route.} \@namedef{PFln@lead@narrower}{Narrower in Lean.} \@namedef{PFln@lead@different}{Different form in Lean.} \@namedef{PFln@lead@conditional}{Conditional in Lean.} \@namedef{PFln@lead@details}{Details in Lean.} \@namedef{PFln@lead@hand}{Not formalized.} \@namedef{PFln@lead@none}{} \newcommand*\PFln@lead[1]{% \ifcsname PFln@lead@#1\endcsname \edef\PFln@tmp{\csname PFln@lead@#1\endcsname}% \ifx\PFln@tmp\@empty\else{\itshape\PFln@tmp}\space\fi \else\PackageWarning{leannote}{Unknown note kind `#1'}\fi} \newcommand*\PFln@map[2]{\@namedef{PFln@map@#1}{#2}} \PFln@map{Same route}{same} \PFln@map{Same route (planned)}{same} \PFln@map{Same statement}{formal} \PFln@map{Same statement, different route}{route} \PFln@map{Same statement, different route (planned)}{route} \PFln@map{Same statement, different route (source-checked)}{route} \PFln@map{Different route (planned)}{route} \PFln@map{Lean scope narrower}{narrower} \PFln@map{Lean scope narrower (planned)}{narrower} \PFln@map{Lean scope narrower (source-checked)}{narrower} \PFln@map{Different form}{different} \PFln@map{Lean scope conditional}{conditional} \PFln@map{Proof by formalization (planned)}{details} \PFln@map{Not formalized}{hand} \PFln@map{Lean endpoint; route not compared}{none} \PFln@map{Lean status not yet assigned}{none} \newcommand\provisionalref[4]{% \ifcsname PFln@map@#3\endcsname \edef\PFln@key{\csname PFln@map@#3\endcsname}% \else\def\PFln@key{none}\fi \expandafter\leannote\expandafter{\PFln@key}{#4}} \newcommand\PFln@needspace[1]{% \ifhmode\par\fi \@tempskipb\lastskip\vskip-\@tempskipb\penalty-100 \begingroup\setlength\dimen@{#1}\advance\dimen@\@tempskipb \dimen@ii\pagegoal\advance\dimen@ii-\pagetotal \ifdim\dimen@>\dimen@ii \PackageInfo{leannote}{page break before note: needs \the\dimen@, has \the\dimen@ii}% \ifdim\dimen@ii>\z@\vfil\fi\break \else\vskip\@tempskipb\fi \endgroup} \newsavebox\PFln@box \newcount\PFln@cnt \newif\ifPFln@armed \newcommand\PFln@build[1]{% \global\setbox\PFln@box\vtop{\hsize\marginparwidth\leannotestyle\leavevmode{\leannotecolor#1}\par}} \newcommand\leannote[2]{\PFln@build{\PFln@lead{#1}#2}\PFln@dispatch} \newcommand\leannoteexplain[1]{\PFln@build{#1}\PFln@dispatch} \newcommand\PFln@dispatch{% \ifvmode \if@nobreak\PackageInfo{leannote}{note after a heading: no room requested}% \else\PFln@needspace{\dimexpr\dp\PFln@box+2\baselineskip\relax}\fi \global\PFln@armedtrue \expandafter\PFln@look \else \PFln@place \fi} \newcommand\PFln@look{\@ifnextchar\begin{}{\@ifnextchar\paragraph\PFln@para\PFln@leave}} \newcommand\PFln@para[2]{#1{#2}\leavevmode\PFln@fire\ignorespaces} \newcommand\PFln@leave{\leavevmode\PFln@fire} \newcommand\PFln@fire{\ifPFln@armed\global\PFln@armedfalse\PFln@place\fi} \let\PFln@orig@dth\deferred@thm@head \def\deferred@thm@head#1{\PFln@orig@dth{\PFln@fire#1}} \newcommand\PFln@place{% \global\advance\PFln@cnt\@ne \zsavepos{PFln\the\PFln@cnt}% \setbox\z@\hbox{\copy\PFln@box}% \PackageInfo{leannote}{note \the\PFln@cnt: depth \the\dp\z@\space height \the\ht\z@}% \immediate\write\@mainaux{\string\PFlnsize{\the\PFln@cnt}{\the\ht\z@}{\the\dp\z@}}% \ht\z@\z@\dp\z@\z@ \rlap{\kern\dimexpr1in+\hoffset+\oddsidemargin+\textwidth+\marginparsep -\zposx{PFln\the\PFln@cnt}sp\relax\box\z@}} \providecommand\PFlnsize[3]{} \AtBeginDocument{\zref@addprop{savepos}{abspage}} \makeatother \endinput \end{filecontents*} \begin{filecontents*}[overwrite]{proofnote-cover.sty} \NeedsTeXFormat{LaTeX2e} \ProvidesPackage{proofnote-cover}[2026/11/17 AI proof note cover helpers (mock-up)] \RequirePackage{xcolor,tikz,array,tabularx,booktabs,enumitem} \newcommand{\tbd}[1]{\tikz[baseline=(n.base)]\node[draw,dashed,line width=0.35pt, inner xsep=1.6pt,inner ysep=0.9pt,rounded corners=0.8pt](n){\normalfont\sffamily\scriptsize\color[gray]{0.35}#1};} \newcommand{\stSame}{\tikz[baseline=-0.6ex]\fill (0,0) circle (0.9mm);} \newcommand{\stNarrow}{\tikz[baseline=-0.6ex]{\draw[line width=0.4pt] (0,0) circle (0.9mm); \fill (0,-0.9mm) arc (-90:90:0.9mm) -- cycle;}} \newcommand{\stHand}{\tikz[baseline=-0.6ex]\draw[line width=0.5pt] (0,0) circle (0.9mm);} \newcommand{\stImp}{\tikz[baseline=-0.6ex]\draw[line width=0.5pt] (-0.8mm,-0.8mm) rectangle (0.8mm,0.8mm);} \newcommand{\pnband}[2]{% \noindent\begin{tikzpicture} \useasboundingbox (0,2mm) rectangle (\linewidth,-6.5mm); \node[anchor=west,inner sep=0pt] at (0,0) {\normalfont\sffamily\bfseries\small\addfontfeature{}AI\ \ PROOF\ \ NOTE}; \node[anchor=east,inner sep=0pt] at (\linewidth,0) {\normalfont\sffamily\small #1}; \draw[line width=1.2pt] (0,-2.6mm) -- (\linewidth,-2.6mm); \node[anchor=west,inner sep=0pt] at (0,-5.4mm) {\normalfont\sffamily\scriptsize #2}; \end{tikzpicture}\par} \providecommand{\addfontfeature}[1]{} \newlength{\pnlab}\setlength{\pnlab}{0.92in} \newlength{\pnsep}\setlength{\pnsep}{0.16in} \newenvironment{pnrow}[1]{% \par\noindent\begin{minipage}[t]{\pnlab}\raggedright\normalfont\sffamily\scriptsize\bfseries \MakeUppercase{#1}\end{minipage}\hspace{\pnsep}% \begin{minipage}[t]{\dimexpr\linewidth-\pnlab-\pnsep\relax}\small}{% \end{minipage}\par} \newcommand{\pnrule}{\par\vspace{0.7ex}\noindent{\color[gray]{0.55}\rule{\linewidth}{0.3pt}}\par\vspace{0.6ex}} \newenvironment{pnthm}[2]{% \par\noindent\textbf{#1}\ \textit{(#2)}\ \ignorespaces}{\par} \newcommand{\pnstatus}[1]{\par\vspace{0.5ex}\noindent{\raggedright\normalfont\sffamily\scriptsize #1\par}} \endinput \end{filecontents*} \documentclass[11pt]{article} \usepackage[T1]{fontenc} \usepackage[type1]{newpxtext} \usepackage{amsmath,amsthm,amssymb,mathtools} \usepackage{newpxmath} \usepackage[letterpaper,textwidth=5.4in,textheight=8.6in,left=0.8in,marginparsep=0.2in,marginparwidth=1.3in]{geometry} \PassOptionsToPackage{noadjust}{marginnote}\usepackage{xcolor,marginnote,microtype,graphicx,caption,booktabs,tabularx,xltabular,needspace,array,enumitem,tikz} \usepackage[super,sort&compress]{natbib} \usepackage[hidelinks,bookmarksnumbered]{hyperref} \hypersetup{pdftitle={On sequential ideal measurements in the exceptional Jordan algebra},pdfauthor={Robert R. MacGregor}} \setlength{\emergencystretch}{2em} \captionsetup{font=small,hypcap=false} \makeatletter\setlength{\@fptop}{0pt}\makeatother \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{computation}[theorem]{Calculation} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{example}[theorem]{Example} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \newenvironment{conceptfigure}{\begin{figure}[htbp]\centering}{\end{figure}} \usepackage{leannote} \usepackage[titles]{tocloft} \renewcommand{\cftdot}{} \setlength{\cftbeforesecskip}{3pt} \renewcommand{\cftsecfont}{\bfseries} \renewcommand{\cftsecpagefont}{\bfseries} \setcounter{tocdepth}{2} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\Ha}{\mathbb{H}} \newcommand{\Oct}{\mathbb{O}} \newcommand{\Q}{\mathbb{Q}} \newcommand{\J}{J} \newcommand{\HR}{H_3(\R)} \newcommand{\HC}{H_3(\C)} \newcommand{\HH}{H_3(\Ha)} \newcommand{\HO}{H_3(\Oct)} \newcommand{\OP}{\Oct P^2} \newcommand{\tr}{\operatorname{tr}} \newcommand{\Tr}{\operatorname{Tr}} \newcommand{\Jord}{\operatorname{Jord}} \newcommand{\jp}{\circ} \newcommand{\Aut}{\operatorname{Aut}} \newcommand{\Spin}{\operatorname{Spin}} \newcommand{\Ffour}{F_4} \newcommand{\Esix}{E_{6(-26)}} \newcommand{\spn}{\operatorname{span}} \newcommand{\ip}[2]{\langle #1,#2\rangle} \newcommand{\Acal}{\mathcal{A}} \newcommand{\Dtet}{\mathcal{D}} \newcommand{\BH}{B(\mathcal{H})} \newcommand{\Hil}{\mathcal{H}} \newcommand{\sa}{\mathrm{sa}} \newcommand{\TV}{\operatorname{TV}} \newcommand{\id}{\mathrm{id}} \newcommand{\CHECK}[1]{\textcolor{orange}{\textsf{[check: #1]}}} \usepackage{proofnote-cover} \newcommand{\leanrepo}{\href{https://github.com/macgregor-sudo/albert-sequential-measurements}{\nolinkurl{github.com/macgregor-sudo/albert-sequential-measurements}}} \newcommand{\companionref}[1]{the companion paper (#1)} \newcommand{\coverhead}[1]{{\normalfont\bfseries\scshape #1}\par\vspace{-0.25\baselineskip}} \usepackage{etoolbox} \expandafter\def\csname PFref@AbramskyBrandenburger2011\endcsname{S.~Abramsky and A.~Brandenburger, The sheaf-theoretic structure of non-locality and contextuality, \emph{New J. Phys.} \textbf{13} (2011) 113036. \href{https://doi.org/10.1088/1367-2630/13/11/113036}{doi:10.1088/1367-2630/13/11/113036}. \href{https://arxiv.org/abs/1102.0264}{arXiv:1102.0264}} \expandafter\def\csname PFref@Albert1934\endcsname{A.~A. Albert, On a certain algebra of quantum mechanics, \emph{Ann. of Math.} (2) \textbf{35} (1934) 65--73. \href{https://doi.org/10.2307/1968118}{doi:10.2307/1968118}} \expandafter\def\csname PFref@AlbertPaige1959\endcsname{A.~A. Albert and L.~J. Paige, On a homomorphism property of certain Jordan algebras, \emph{Trans. Amer. Math. Soc.} \textbf{93} (1959) 20--29. \href{https://doi.org/10.1090/S0002-9947-1959-0108524-2}{doi:10.1090/S0002-9947-1959-0108524-2}} \expandafter\def\csname PFref@AlfsenShultz2003\endcsname{E.~M. Alfsen and F.~W. Shultz, \emph{Geometry of State Spaces of Operator Algebras}, Birkh\"auser, Boston, 2003} \expandafter\def\csname PFref@AlfsenShultzStormer1978\endcsname{E.~M. Alfsen, F.~W. Shultz and E.~St{\o}rmer, A Gelfand--Neumark theorem for Jordan algebras, \emph{Adv. Math.} \textbf{28} (1978) 11--56. \href{https://doi.org/10.1016/0001-8708(78)90044-0}{doi:10.1016/0001-8708(78)90044-0}} \expandafter\def\csname PFref@Araki1980b\endcsname{H.~Araki, On a characterization of the state space of quantum mechanics, \emph{Commun. Math. Phys.} \textbf{75} (1980) 1--24. \href{https://doi.org/10.1007/BF01962588}{doi:10.1007/BF01962588}} \expandafter\def\csname PFref@AsoreyScolariciSolombrino2007\endcsname{M.~Asorey, G.~Scolarici and L.~Solombrino, Complex projections of completely positive quaternionic maps, \emph{Theoret. and Math. Phys.} \textbf{151} (2007) 735--743. \href{https://www.mathnet.ru/eng/tmf6051}{Math-Net.Ru}} \expandafter\def\csname PFref@Baez2002\endcsname{J.~C. Baez, The octonions, \emph{Bull. Amer. Math. Soc.} \textbf{39} (2002) 145--205. \href{https://doi.org/10.1090/S0273-0979-01-00934-X}{doi:10.1090/S0273-0979-01-00934-X}. \href{https://arxiv.org/abs/math/0105155}{arXiv:math/0105155}} \expandafter\def\csname PFref@Baez2012\endcsname{J.~C. Baez, Division algebras and quantum theory, \emph{Found. Phys.} \textbf{42} (2012) 819--855. \href{https://doi.org/10.1007/s10701-011-9566-z}{doi:10.1007/s10701-011-9566-z}. \href{https://arxiv.org/abs/1101.5690}{arXiv:1101.5690}} \expandafter\def\csname PFref@BaezBokorBoyle2026\endcsname{J.~C. Baez, E.~Bokor and L.~Boyle, Jordan pair quantum theory and the standard model, \href{https://arxiv.org/abs/2607.10833}{arXiv:2607.10833} (2026)} \expandafter\def\csname PFref@BaezSchwahn2026\endcsname{J.~C. Baez and P.~Schwahn, The Standard Model gauge group from the exceptional Jordan algebra, \href{https://arxiv.org/abs/2606.15235}{arXiv:2606.15235} (2026)} \expandafter\def\csname PFref@BarnumHilgert2020\endcsname{H.~Barnum and J.~Hilgert, Spectral properties of convex bodies, \emph{J. Lie Theory} \textbf{30} (2020) 315--344. \href{https://doi.org/10.5802/jolt.1118}{doi:10.5802/jolt.1118}. \href{https://arxiv.org/abs/1904.03753}{arXiv:1904.03753} (preprint title: Strongly symmetric spectral convex bodies are Jordan algebra state spaces)} \expandafter\def\csname PFref@BarnumWilce2014\endcsname{H.~Barnum and A.~Wilce, Local tomography and the Jordan structure of quantum theory, \emph{Found. Phys.} \textbf{44} (2014) 192--212. \href{https://doi.org/10.1007/s10701-014-9777-1}{doi:10.1007/s10701-014-9777-1}. \href{https://arxiv.org/abs/1202.4513}{arXiv:1202.4513}} \expandafter\def\csname PFref@BGW2015\endcsname{H.~Barnum, M.~A. Graydon and A.~Wilce, Some nearly quantum theories, in \emph{Proc. 12th Int. Workshop on Quantum Physics and Logic (QPL 2015)}, Electron. Proc. Theor. Comput. Sci. \textbf{195} (2015) 59--70. \href{https://doi.org/10.4204/EPTCS.195.5}{doi:10.4204/EPTCS.195.5}. \href{https://arxiv.org/abs/1507.06278}{arXiv:1507.06278}} \expandafter\def\csname PFref@BGW2020\endcsname{H.~Barnum, M.~A. Graydon and A.~Wilce, Composites and categories of Euclidean Jordan algebras, \emph{Quantum} \textbf{4} (2020) 359. \href{https://doi.org/10.22331/q-2020-11-08-359}{doi:10.22331/q-2020-11-08-359}. \href{https://arxiv.org/abs/1606.09331}{arXiv:1606.09331}} \expandafter\def\csname PFref@BMU2014\endcsname{H.~Barnum, M.~P. M\"uller and C.~Ududec, Higher-order interference and single-system postulates characterizing quantum theory, \emph{New J. Phys.} \textbf{16} (2014) 123029. \href{https://doi.org/10.1088/1367-2630/16/12/123029}{doi:10.1088/1367-2630/16/12/123029}. \href{https://arxiv.org/abs/1403.4147}{arXiv:1403.4147}} \expandafter\def\csname PFref@Boyle2020\endcsname{L.~Boyle, The standard model, the exceptional Jordan algebra, and triality, \emph{J. Math. Phys.} \textbf{67} (2026) 071701. \href{https://doi.org/10.1063/5.0336175}{doi:10.1063/5.0336175}. \href{https://arxiv.org/abs/2006.16265}{arXiv:2006.16265}} \expandafter\def\csname PFref@BoyleFarnsworth2020\endcsname{L.~Boyle and S.~Farnsworth, The standard model, the Pati--Salam model, and ``Jordan geometry'', \emph{New J. Phys.} \textbf{22} (2020) 073023. \href{https://doi.org/10.1088/1367-2630/ab9709}{doi:10.1088/1367-2630/ab9709}. \href{https://arxiv.org/abs/1910.11888}{arXiv:1910.11888}} \expandafter\def\csname PFref@Brehm1990\endcsname{U.~Brehm, The shape invariant of triangles and trigonometry in two-point homogeneous spaces, \emph{Geom. Dedicata} \textbf{33} (1990) 59--76. \href{https://doi.org/10.1007/BF00147601}{doi:10.1007/BF00147601}} \expandafter\def\csname PFref@BMKG2013\endcsname{C.~Budroni, T.~Moroder, M.~Kleinmann and O.~G\"uhne, Bounding temporal quantum correlations, \emph{Phys. Rev. Lett.} \textbf{111} (2013) 020403. \href{https://doi.org/10.1103/PhysRevLett.111.020403}{doi:10.1103/PhysRevLett.111.020403}. \href{https://arxiv.org/abs/1302.6223}{arXiv:1302.6223}} \expandafter\def\csname PFref@BudroniEmary2014\endcsname{C.~Budroni and C.~Emary, Temporal quantum correlations and Leggett--Garg inequalities in multilevel systems, \emph{Phys. Rev. Lett.} \textbf{113} (2014) 050401. \href{https://doi.org/10.1103/PhysRevLett.113.050401}{doi:10.1103/PhysRevLett.113.050401}. \href{https://arxiv.org/abs/1309.3678}{arXiv:1309.3678}} \expandafter\def\csname PFref@BudroniFagundesKleinmann2019\endcsname{C.~Budroni, G.~Fagundes and M.~Kleinmann, Memory cost of temporal correlations, \emph{New J. Phys.} \textbf{21} (2019) 093018. \href{https://doi.org/10.1088/1367-2630/ab3cb4}{doi:10.1088/1367-2630/ab3cb4}. \href{https://arxiv.org/abs/1902.06517}{arXiv:1902.06517}} \expandafter\def\csname PFref@BuschGrabowskiLahti1995\endcsname{P.~Busch, M.~Grabowski and P.~J. Lahti, Repeatable measurements in quantum theory: their role and feasibility, \emph{Found. Phys.} \textbf{25} (1995) 1239--1266. \href{https://doi.org/10.1007/BF02055331}{doi:10.1007/BF02055331}} \expandafter\def\csname PFref@BuschSingh1998\endcsname{P.~Busch and J.~Singh, L\"uders theorem for unsharp quantum measurements, \emph{Phys. Lett. A} \textbf{249} (1998) 10--12. \href{https://doi.org/10.1016/S0375-9601(98)00704-X}{doi:10.1016/S0375-9601(98)00704-X}. \href{https://arxiv.org/abs/1304.0054}{arXiv:1304.0054}} \expandafter\def\csname PFref@Catto2003\endcsname{S.~Catto, Exceptional projective geometries and internal symmetries, \href{https://arxiv.org/abs/hep-th/0302079}{arXiv:hep-th/0302079} (2003)} \expandafter\def\csname PFref@ChiribellaYuan2014\endcsname{G.~Chiribella and X.~Yuan, Measurement sharpness cuts nonlocality and contextuality in every physical theory, \href{https://arxiv.org/abs/1404.3348}{arXiv:1404.3348} (2014)} \expandafter\def\csname PFref@ChiribellaYuan2018\endcsname{G.~Chiribella and X.~Yuan, Bridging the gap between general probabilistic theories and the device-independent framework for nonlocality and contextuality, \emph{Inf. Comput.} \textbf{250} (2016) 15--49. \href{https://doi.org/10.1016/j.ic.2016.02.006}{doi:10.1016/j.ic.2016.02.006}. \href{https://arxiv.org/abs/1504.02395}{arXiv:1504.02395}} \expandafter\def\csname PFref@Cho2019\endcsname{K.~Cho, \emph{Effectuses in Categorical Quantum Foundations}, PhD thesis, Radboud University Nijmegen, 2019. \href{https://arxiv.org/abs/1910.12198}{arXiv:1910.12198}} \expandafter\def\csname PFref@ClementeKofler2016\endcsname{L.~Clemente and J.~Kofler, No fine theorem for macrorealism: limitations of the Leggett--Garg inequality, \emph{Phys. Rev. Lett.} \textbf{116} (2016) 150401. \href{https://arxiv.org/abs/1509.00348}{arXiv:1509.00348}} \expandafter\def\csname PFref@DaviesLewis1970\endcsname{E.~B. Davies and J.~T. Lewis, An operational approach to quantum probability, \emph{Commun. Math. Phys.} \textbf{17} (1970) 239--260} \expandafter\def\csname PFref@DuboisViolette2016\endcsname{M.~Dubois-Violette, Exceptional quantum geometry and particle physics, \emph{Nucl. Phys. B} \textbf{912} (2016) 426--449. \href{https://doi.org/10.1016/j.nuclphysb.2016.04.018}{doi:10.1016/j.nuclphysb.2016.04.018}. \href{https://arxiv.org/abs/1604.01247}{arXiv:1604.01247}} \expandafter\def\csname PFref@DuboisVioletteTodorov2018\endcsname{M.~Dubois-Violette and I.~Todorov, Exceptional quantum geometry and particle physics~II, \emph{Nucl. Phys. B} \textbf{938} (2019) 751--761. \href{https://doi.org/10.1016/j.nuclphysb.2018.12.012}{doi:10.1016/j.nuclphysb.2018.12.012}. \href{https://arxiv.org/abs/1808.08110}{arXiv:1808.08110}} \expandafter\def\csname PFref@Ehrlich2026\endcsname{B.~Ehrlich, \texttt{sequential-products-lean}: classification of sequential products on Hermitian matrix algebras (Lean formalisation), repository \texttt{ehrlich-b/sequential-products-lean}, 2026; Palomar registry entries PALOMAR-2026-08-21-000009 and PALOMAR-2026-08-22-000006 (commit \texttt{9c6d92f5}), Apache License~2.0} \expandafter\def\csname PFref@Ehrlich2026ca\endcsname{B.~Ehrlich, \texttt{composition-algebras}: Euclidean composition algebras over $\mathbb R$ in Lean~4, repository \texttt{ehrlich-b/composition-algebras}, 2026; Palomar registry entries PALOMAR-2026-08-29-000015 (commit \texttt{e37e22b0}) and PALOMAR-2026-08-29-000018 (commit \texttt{98d4437c}), Apache License~2.0} \expandafter\def\csname PFref@Ehrlich2026ej\endcsname{B.~Ehrlich, \texttt{euclidean-jordan}: Euclidean Jordan algebras in Lean~4, repository \texttt{ehrlich-b/euclidean-jordan}, 2026; Palomar registry entries PALOMAR-2026-08-23-000005, PALOMAR-2026-08-23-000006, PALOMAR-2026-08-29-000001 and PALOMAR-2026-08-29-000008, Apache License~2.0} \expandafter\def\csname PFref@ElduqueIltyakov1999\endcsname{A.~Elduque and A.~V. Iltyakov, On polynomial invariants of exceptional simple algebraic groups, \emph{Canad. J. Math.} \textbf{51} (1999) 506--522. \href{https://doi.org/10.4153/CJM-1999-023-2}{doi:10.4153/CJM-1999-023-2}} \expandafter\def\csname PFref@FNW1992\endcsname{M.~Fannes, B.~Nachtergaele and R.~F. Werner, Finitely correlated states on quantum spin chains, \emph{Commun. Math. Phys.} \textbf{144} (1992) 443--490. \href{https://doi.org/10.1007/BF02099178}{doi:10.1007/BF02099178}} \expandafter\def\csname PFref@FLW2024\endcsname{M.~Fanizza, J.~Lumbreras and A.~Winter, Quantum theory in finite dimension cannot explain every general process with finite memory, \emph{Commun. Math. Phys.} \textbf{405} (2024) 50. \href{https://doi.org/10.1007/s00220-023-04913-4}{doi:10.1007/s00220-023-04913-4}. \href{https://arxiv.org/abs/2209.11225}{arXiv:2209.11225}. (Journal Theorems~3--4 are Theorems~7--8 of arXiv v2.)} \expandafter\def\csname PFref@FarautKoranyi1994\endcsname{J.~Faraut and A.~Kor\'anyi, \emph{Analysis on Symmetric Cones}, Oxford Mathematical Monographs, The Clarendon Press, Oxford University Press, New York, 1994} \expandafter\def\csname PFref@FernandesWagnerNovoGalvao2024\endcsname{C.~Fernandes, R.~Wagner, L.~Novo and E.~F. Galv\~ao, Unitary-invariant witnesses of quantum imaginarity, \emph{Phys. Rev. Lett.} \textbf{133} (2024) 190201. \href{https://arxiv.org/abs/2403.15066}{arXiv:2403.15066}} \expandafter\def\csname PFref@GarnerMuller2020\endcsname{A.~J.~P. Garner and M.~P. M\"uller, Characterization of the probabilistic models that can be embedded in quantum theory, \href{https://arxiv.org/abs/2004.06136}{arXiv:2004.06136} (2020)} \expandafter\def\csname PFref@Glennie1966\endcsname{C.~M. Glennie, Some identities valid in special Jordan algebras but not valid in all Jordan algebras, \emph{Pacific J. Math.} \textbf{16} (1966) 47--59. \href{https://doi.org/10.2140/pjm.1966.16.47}{doi:10.2140/pjm.1966.16.47}} \expandafter\def\csname PFref@GMS2025\endcsname{I.~Gorshkov, A.~Mamontov and A.~Staroletov, On Jordan algebras that are factors of Matsuo algebras, \emph{Int. Electron. J. Algebra} \textbf{37} (2025) 70--90. \href{https://doi.org/10.24330/ieja.1488467}{doi:10.24330/ieja.1488467}. \href{https://arxiv.org/abs/2305.10958}{arXiv:2305.10958}} \expandafter\def\csname PFref@GorshkovStaroletov2020\endcsname{I.~Gorshkov and A.~Staroletov, On primitive 3-generated axial algebras of Jordan type, \emph{J. Algebra} \textbf{563} (2020) 74--99. \href{https://doi.org/10.1016/j.jalgebra.2020.07.014}{doi:10.1016/j.jalgebra.2020.07.014}. \href{https://arxiv.org/abs/2005.13791}{arXiv:2005.13791}} \expandafter\def\csname PFref@GudderGreechie2002\endcsname{S.~Gudder and R.~Greechie, Sequential products on effect algebras, \emph{Rep. Math. Phys.} \textbf{49} (2002) 87--111. \href{https://doi.org/10.1016/S0034-4877(02)80007-6}{doi:10.1016/S0034-4877(02)80007-6}} \expandafter\def\csname PFref@GPR1978\endcsname{M.~G\"unaydin, C.~Piron and H.~Ruegg, Moufang plane and octonionic quantum mechanics, \emph{Commun. Math. Phys.} \textbf{61} (1978) 69--85. \href{https://doi.org/10.1007/BF01609468}{doi:10.1007/BF01609468}} \expandafter\def\csname PFref@GunaydinGursey1973\endcsname{M.~G\"unaydin and F.~G\"ursey, Quark structure and octonions, \emph{J. Math. Phys.} \textbf{14} (1973) 1651--1667. \href{https://doi.org/10.1063/1.1666240}{doi:10.1063/1.1666240}} \expandafter\def\csname PFref@HarveyLawson1982\endcsname{R.~Harvey and H.~B. Lawson, Jr., Calibrated geometries, \emph{Acta Math.} \textbf{148} (1982) 47--157. \href{https://doi.org/10.1007/BF02392726}{doi:10.1007/BF02392726}} \expandafter\def\csname PFref@HeinosaariReitznerStano2008\endcsname{T.~Heinosaari, D.~Reitzner and P.~Stano, Notes on joint measurability of quantum observables, \emph{Found. Phys.} \textbf{38} (2008) 1133--1147. \href{https://doi.org/10.1007/s10701-008-9256-7}{doi:10.1007/s10701-008-9256-7}. \href{https://arxiv.org/abs/0811.0783}{arXiv:0811.0783}} \expandafter\def\csname PFref@HSGB2018\endcsname{J.~Hoffmann, C.~Spee, O.~G\"uhne and C.~Budroni, Structure of temporal correlations of a qubit, \emph{New J. Phys.} \textbf{20} (2018) 102001. \href{https://arxiv.org/abs/1712.01234}{arXiv:1712.01234}} \expandafter\def\csname PFref@Hoggar1977\endcsname{S.~G. Hoggar, New sets of equi-isoclinic $n$-planes from old, \emph{Proc. Edinburgh Math. Soc.} \textbf{20} (1977) 287--291. \href{https://doi.org/10.1017/S0013091500026511}{doi:10.1017/S0013091500026511}} \expandafter\def\csname PFref@Hsiang1989\endcsname{W.-Y. Hsiang, On the laws of trigonometries of two-point homogeneous spaces, \emph{Ann. Global Anal. Geom.} \textbf{7} (1989) 29--45. \href{https://doi.org/10.1007/BF00137400}{doi:10.1007/BF00137400}} \expandafter\def\csname PFref@Iltyakov1997\endcsname{A.~V. Iltyakov, On generators of the algebra of polynomial invariants of group $F_4$, University of Sydney Research Report 97-4, 1997} \expandafter\def\csname PFref@JordanvNWigner1934\endcsname{P.~Jordan, J.~von Neumann and E.~Wigner, On an algebraic generalization of the quantum mechanical formalism, \emph{Ann. of Math.} \textbf{35} (1934) 29--64. \href{https://doi.org/10.2307/1968117}{doi:10.2307/1968117}} \expandafter\def\csname PFref@KhandelwalTavakoli2025\endcsname{S.~Khandelwal and A.~Tavakoli, Simulating quantum instruments with projective measurements and quantum post-processing, \emph{Phys. Rev. Lett.} \textbf{135} (2025) 040202. \href{https://arxiv.org/abs/2503.00956}{arXiv:2503.00956}} \expandafter\def\csname PFref@Kleinmann2014\endcsname{M.~Kleinmann, Sequences of projective measurements in generalized probabilistic models, \emph{J. Phys. A} \textbf{47} (2014) 455304. \href{https://doi.org/10.1088/1751-8113/47/45/455304}{doi:10.1088/1751-8113/47/45/455304}. \href{https://arxiv.org/abs/1402.3583}{arXiv:1402.3583}} \expandafter\def\csname PFref@Kollross\endcsname{A.~Kollross, Octonions, triality, the exceptional Lie algebra $F_4$ and polar actions on the Cayley hyperbolic plane, \emph{Internat. J. Math.} \textbf{31} (2020) 2050051. \href{https://doi.org/10.1142/S0129167X20500512}{doi:10.1142/S0129167X20500512}. \href{https://arxiv.org/abs/1802.08075}{arXiv:1802.08075}} \expandafter\def\csname PFref@Krasnov2019\endcsname{K.~Krasnov, SO(9) characterization of the Standard Model gauge group, \emph{J. Math. Phys.} \textbf{62} (2021) 021703. \href{https://doi.org/10.1063/5.0039941}{doi:10.1063/5.0039941}. \href{https://arxiv.org/abs/1912.11282}{arXiv:1912.11282}} \expandafter\def\csname PFref@Luders1951\endcsname{G.~L\"uders, \"Uber die Zustands\"anderung durch den Me{\ss}proze{\ss}, \emph{Ann. Phys. (Leipzig)} (6) \textbf{8} (1950) 322--328. \href{https://doi.org/10.1002/andp.19504430510}{doi:10.1002/andp.19504430510}; English translation: Concerning the state-change due to the measurement process, \emph{Ann. Phys. (Leipzig)} \textbf{15} (2006) 663--670, \href{https://doi.org/10.1002/andp.20065180904}{doi:10.1002/andp.20065180904}} \expandafter\def\csname PFref@Physlib\endcsname{T.~O. Diem, in Physlib (\texttt{leanprover-community/physlib}), Lean library, commit \texttt{08a6c1b5}, files \nolinkurl{PhyslibAlpha/ProbabilisticTheory/JordanOrderUnit/Quadratic/Fundamental.lean}, \nolinkurl{.../Power/Associative.lean}, \nolinkurl{.../Operator.lean} and \nolinkurl{.../Quadratic/Triple.lean}, copyright 2026 Tom Ole Diem, Apache License~2.0. \url{https://github.com/leanprover-community/physlib}} \expandafter\def\csname PFref@MacGregorPhi2026\endcsname{R.~R. MacGregor, Finite observations and recognition of the compact Albert algebra, Palomar registry entry PALOMAR-2026-10-16-000013, 16 October 2026, Lean repository \texttt{macgregor-sudo/compact-albert-recognition}, commit \texttt{40d3b34}} \expandafter\def\csname PFref@MasilloScolariciSozzo2009\endcsname{F.~Masillo, G.~Scolarici and S.~Sozzo, Proper versus improper mixtures: Toward a quaternionic quantum mechanics, \emph{Theoret. and Math. Phys.} \textbf{160} (2009) 1006--1013. \href{https://arxiv.org/abs/0901.0795}{arXiv:0901.0795}} \expandafter\def\csname PFref@McCrimmon2004\endcsname{K.~McCrimmon, \emph{A Taste of Jordan Algebras}, Universitext, Springer, New York, 2004} \expandafter\def\csname PFref@McCrimmonIdentities2004\endcsname{K.~McCrimmon, The role of identities in Jordan algebras, \emph{Resenhas IME-USP} \textbf{6} (2004), no.~2/3, 265--280. \href{https://doi.org/10.11606/resimeusp.v6i2/3.75294}{doi:10.11606/resimeusp.v6i2/3.75294}} \expandafter\def\csname PFref@MonrasWinter2016\endcsname{A.~Monr\`as and A.~Winter, Quantum learning of classical stochastic processes: the completely-positive realization problem, \emph{J. Math. Phys.} \textbf{57} (2016) 015219. \href{https://doi.org/10.1063/1.4936935}{doi:10.1063/1.4936935}. \href{https://arxiv.org/abs/1412.3634}{arXiv:1412.3634}} \expandafter\def\csname PFref@MullerGarner2023\endcsname{M.~P. M\"uller and A.~J.~P. Garner, Testing quantum theory by generalizing noncontextuality, \emph{Phys. Rev. X} \textbf{13} (2023) 041001. \href{https://doi.org/10.1103/PhysRevX.13.041001}{doi:10.1103/PhysRevX.13.041001}. \href{https://arxiv.org/abs/2112.09719}{arXiv:2112.09719}} \expandafter\def\csname PFref@NPA2008\endcsname{M.~Navascu\'es, S.~Pironio and A.~Ac\'in, A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations, \emph{New J. Phys.} \textbf{10} (2008) 073013. \href{https://doi.org/10.1088/1367-2630/10/7/073013}{doi:10.1088/1367-2630/10/7/073013}. \href{https://arxiv.org/abs/0803.4290}{arXiv:0803.4290}} \expandafter\def\csname PFref@Niestegge2008\endcsname{G.~Niestegge, A representation of quantum measurement in order-unit spaces, \emph{Found. Phys.} \textbf{38} (2008) 783--795. \href{https://doi.org/10.1007/s10701-008-9236-y}{doi:10.1007/s10701-008-9236-y}. \href{https://arxiv.org/abs/1001.3633}{arXiv:1001.3633}} \expandafter\def\csname PFref@Niestegge2012\endcsname{G.~Niestegge, Conditional probability, three-slit experiments, and the Jordan algebra structure of quantum mechanics, \emph{Adv. Math. Phys.} \textbf{2012} (2012) 156573. \href{https://doi.org/10.1155/2012/156573}{doi:10.1155/2012/156573}. \href{https://arxiv.org/abs/0912.0203}{arXiv:0912.0203}} \expandafter\def\csname PFref@Niestegge2020a\endcsname{G.~Niestegge, A simple and quantum-mechanically motivated characterization of the formally real Jordan algebras, \emph{Proc. R. Soc. A} \textbf{476} (2020) 20190604. \href{https://doi.org/10.1098/rspa.2019.0604}{doi:10.1098/rspa.2019.0604}. \href{https://arxiv.org/abs/1905.04189}{arXiv:1905.04189}} \expandafter\def\csname PFref@Niestegge2020b\endcsname{G.~Niestegge, Quantum probability's algebraic origin, \emph{Entropy} \textbf{22} (2020) 1196. \href{https://doi.org/10.3390/e22111196}{doi:10.3390/e22111196}. \href{https://arxiv.org/abs/2009.08489}{arXiv:2009.08489}} \expandafter\def\csname PFref@Niestegge2026a\endcsname{G.~Niestegge, On the homogeneity of the quantum transition probability, \href{https://arxiv.org/abs/2601.01936}{arXiv:2601.01936} (2026)} \expandafter\def\csname PFref@Nasmith2022\endcsname{B.~Nasmith, Octonions and the two strictly projective tight 5-designs, \emph{Algebr. Comb.} \textbf{5} (2022) 401--411. \href{https://doi.org/10.5802/alco.215}{doi:10.5802/alco.215}} \expandafter\def\csname PFref@Nasmith2022integers\endcsname{B.~Nasmith, Octonion integers and tight 5-designs, \href{https://arxiv.org/abs/2209.02398v1}{arXiv:2209.02398v1} (2022)} \expandafter\def\csname PFref@Hoehn2026lattice\endcsname{G.~H\"ohn, The unique extremal threemodular lattice of rank 26, the generalized hexagon (2,8), and the tight Cayley-plane 5-design, \href{https://arxiv.org/abs/2609.25086v1}{arXiv:2609.25086v1} (2026)} \expandafter\def\csname PFref@Hoehn2026albert\endcsname{G.~H\"ohn, Definite integral Albert algebras: arithmetic and finite geometry, \href{https://arxiv.org/abs/2609.27550v1}{arXiv:2609.27550v1} (2026)} \expandafter\def\csname PFref@Niestegge2026b\endcsname{G.~Niestegge, The bare necessities of a physically reasonable mathematical model for quantum theory, \href{https://arxiv.org/abs/2607.24474}{arXiv:2607.24474} (2026)} \expandafter\def\csname PFref@OrtegaSantander2002\endcsname{R.~Ortega and M.~Santander, Trigonometry of `complex Hermitian' type homogeneous symmetric spaces, \emph{J. Phys. A} \textbf{35} (2002) 7877--7918. \href{https://arxiv.org/abs/math-ph/0112030}{arXiv:math-ph/0112030}} \expandafter\def\csname PFref@OszmaniecBrodGalvao2021\endcsname{M.~Oszmaniec, D.~J. Brod and E.~F. Galv\~ao, Measuring relational information between quantum states, and applications, \emph{New J. Phys.} \textbf{26} (2024) 013053. \href{https://doi.org/10.1088/1367-2630/ad1a27}{doi:10.1088/1367-2630/ad1a27}. \href{https://arxiv.org/abs/2109.10006}{arXiv:2109.10006}} \expandafter\def\csname PFref@MMG2009\endcsname{M.~McKague, M.~Mosca and N.~Gisin, Simulating quantum systems using real Hilbert spaces, \emph{Phys. Rev. Lett.} \textbf{102} (2009) 020505. \href{https://doi.org/10.1103/PhysRevLett.102.020505}{doi:10.1103/PhysRevLett.102.020505}. \href{https://arxiv.org/abs/0810.1923}{arXiv:0810.1923}} \expandafter\def\csname PFref@Ozawa1984\endcsname{M.~Ozawa, Quantum measuring processes of continuous observables, \emph{J. Math. Phys.} \textbf{25} (1984) 79--87} \expandafter\def\csname PFref@Renou2021\endcsname{M.-O. Renou, D.~Trillo, M.~Weilenmann, T.~P. Le, A.~Tavakoli, N.~Gisin, A.~Ac\'{\i}n and M.~Navascu\'es, Quantum theory based on real numbers can be experimentally falsified, \emph{Nature} \textbf{600} (2021) 625--629. \href{https://doi.org/10.1038/s41586-021-04160-4}{doi:10.1038/s41586-021-04160-4}. \href{https://arxiv.org/abs/2101.10873}{arXiv:2101.10873}} \expandafter\def\csname PFref@Peres1979\endcsname{A.~Peres, Proposed test for complex versus quaternion quantum theory, \emph{Phys. Rev. Lett.} \textbf{42} (1979) 683--686} \expandafter\def\csname PFref@PinedaRua2026\endcsname{S.~Pineda Montoya and J.~H. Rua Munoz, Operational foundations for quaternionic and octonionic quantum models: exact quaternionic channels and error correction, para-linear operators, and categorical closure boundaries beyond associativity, \href{https://arxiv.org/abs/2608.20259}{arXiv:2608.20259} (2026)} \expandafter\def\csname PFref@PNA2010\endcsname{S.~Pironio, M.~Navascu\'es and A.~Ac\'in, Convergent relaxations of polynomial optimization problems with non-commuting variables, \emph{SIAM J. Optim.} \textbf{20} (2010) 2157--2180. \href{https://doi.org/10.1137/090760155}{doi:10.1137/090760155}. \href{https://arxiv.org/abs/0903.4368}{arXiv:0903.4368}} \expandafter\def\csname PFref@PopovaSamoilenko2006\endcsname{N.~D. Popova and Yu.~S. Samoilenko, On the existence of configurations of subspaces in a Hilbert space with fixed angles, \emph{SIGMA} \textbf{2} (2006) 055. \href{https://doi.org/10.3842/SIGMA.2006.055}{doi:10.3842/SIGMA.2006.055}. \href{https://arxiv.org/abs/math/0605717}{arXiv:math/0605717}} \expandafter\def\csname PFref@Racine1977\endcsname{M.~L. Racine, Maximal subalgebras of exceptional Jordan algebras, \emph{J. Algebra} \textbf{46} (1977) 12--21. \href{https://doi.org/10.1016/0021-8693(77)90391-X}{doi:10.1016/0021-8693(77)90391-X}} \expandafter\def\csname PFref@RoelandsTiersma2026\endcsname{M.~Roelands and S.~Tiersma, On the Shirshov--Cohn theorem for JB-algebras, \href{https://arxiv.org/abs/2605.24427v1}{arXiv:2605.24427v1} (2026)} \expandafter\def\csname PFref@SahaSantosAugusiak2020\endcsname{D.~Saha, R.~Santos and R.~Augusiak, Sum-of-squares decompositions for a family of noncontextuality inequalities and self-testing of quantum devices, \emph{Quantum} \textbf{4} (2020) 302. \href{https://doi.org/10.22331/q-2020-08-03-302}{doi:10.22331/q-2020-08-03-302}. \href{https://arxiv.org/abs/2002.12216}{arXiv:2002.12216}} \expandafter\def\csname PFref@SalamonWalpuski2017\endcsname{D.~A. Salamon and T.~Walpuski, Notes on the octonions, in \emph{Proceedings of the 23rd G\"okova Geometry--Topology Conference} (2017) 1--85. \href{https://arxiv.org/abs/1005.2820v4}{arXiv:1005.2820v4}} \expandafter\def\csname PFref@SamoilenkoYershova2006\endcsname{Yu.~S. Samoilenko and Yu.~Yu. Yershova, On equiangular configurations of subspaces of a Hilbert space, \emph{Methods Funct. Anal. Topology} \textbf{12} (2006) 84--96} \expandafter\def\csname PFref@SlomczynskiSzczepanek2018\endcsname{W.~S{\l}omczy\'nski and A.~Szczepanek, Quantum dynamical entropy, chaotic unitaries and complex Hadamard matrices, \emph{IEEE Trans. Inf. Theory} \textbf{63} (2017) 7821--7831. \href{https://doi.org/10.1109/TIT.2017.2751507}{doi:10.1109/TIT.2017.2751507}. \href{https://arxiv.org/abs/1612.03363}{arXiv:1612.03363}} \expandafter\def\csname PFref@Soares2026\endcsname{M.~J. Soares, The Albert algebra in Lean~4: Jordan identity and non-speciality, version~1.0.0, Zenodo, 22 July 2026, repository \texttt{markkasaurus/albert-algebra-lean}, Apache License~2.0. \href{https://doi.org/10.5281/zenodo.21482444}{doi:10.5281/zenodo.21482444}} \expandafter\def\csname PFref@Sorkin1994\endcsname{R.~D. Sorkin, Quantum mechanics as quantum measure theory, \emph{Mod. Phys. Lett. A} \textbf{9} (1994) 3119. \href{https://arxiv.org/abs/gr-qc/9401003}{arXiv:gr-qc/9401003}} \expandafter\def\csname PFref@SpringerVeldkamp2000\endcsname{T.~A. Springer and F.~D. Veldkamp, \emph{Octonions, Jordan Algebras and Exceptional Groups}, Springer Monographs in Mathematics, Springer, Berlin, 2000. \href{https://doi.org/10.1007/978-3-662-12622-6}{doi:10.1007/978-3-662-12622-6}} \expandafter\def\csname PFref@SupicBowles2020\endcsname{I.~\v{S}upi\'c and J.~Bowles, Self-testing of quantum systems: a review, \emph{Quantum} \textbf{4} (2020) 337. \href{https://doi.org/10.22331/q-2020-09-30-337}{doi:10.22331/q-2020-09-30-337}. \href{https://arxiv.org/abs/1904.10042}{arXiv:1904.10042}} \expandafter\def\csname PFref@UdudecBarnumEmerson2011\endcsname{C.~Ududec, H.~Barnum and J.~Emerson, Three slit experiments and the structure of quantum theory, \emph{Found. Phys.} \textbf{41} (2011) 396--405. \href{https://doi.org/10.1007/s10701-010-9429-z}{doi:10.1007/s10701-010-9429-z}. \href{https://arxiv.org/abs/0909.4787}{arXiv:0909.4787}} \expandafter\def\csname PFref@vandeWetering\endcsname{J.~van de Wetering, Three characterisations of the sequential product, \emph{J. Math. Phys.} \textbf{59} (2018) 082202. \href{https://doi.org/10.1063/1.5031089}{doi:10.1063/1.5031089}. \href{https://arxiv.org/abs/1803.08453}{arXiv:1803.08453}} \expandafter\def\csname PFref@vandeWetering2019a\endcsname{J.~van de Wetering, Sequential product spaces are Jordan algebras, \emph{J. Math. Phys.} \textbf{60} (2019) 062201. \href{https://doi.org/10.1063/1.5093504}{doi:10.1063/1.5093504}. \href{https://arxiv.org/abs/1803.11139}{arXiv:1803.11139}} \expandafter\def\csname PFref@VitaglianoBudroni2022\endcsname{G.~Vitagliano and C.~Budroni, Leggett--Garg macrorealism and temporal correlations, \emph{Phys. Rev. A} \textbf{107} (2023) 040101. \href{https://doi.org/10.1103/PhysRevA.107.040101}{doi:10.1103/PhysRevA.107.040101}. \href{https://arxiv.org/abs/2212.11616}{arXiv:2212.11616}} \expandafter\def\csname PFref@Vorobev1962\endcsname{N.~N. Vorob'ev, Consistent families of measures and their extensions, \emph{Theory Probab. Appl.} \textbf{7} (1962) 147--163. \href{https://doi.org/10.1137/1107014}{doi:10.1137/1107014}} \expandafter\def\csname PFref@WWvdW2022\endcsname{B.~Westerbaan and J.~van de Wetering, A computer scientist's reconstruction of quantum theory, \emph{J. Phys. A} \textbf{55} (2022) 384002. \href{https://doi.org/10.1088/1751-8121/ac8459}{doi:10.1088/1751-8121/ac8459}. \href{https://arxiv.org/abs/2109.10707}{arXiv:2109.10707}} \expandafter\def\csname PFref@WWvdW2020\endcsname{A.~Westerbaan, B.~Westerbaan and J.~van de Wetering, The three types of normal sequential effect algebras, \emph{Quantum} \textbf{4} (2020) 378. \href{https://doi.org/10.22331/q-2020-12-24-378}{doi:10.22331/q-2020-12-24-378}. \href{https://arxiv.org/abs/2004.12749}{arXiv:2004.12749}} \expandafter\def\csname PFref@Wigner1963\endcsname{E.~P. Wigner, The problem of measurement, \emph{Am. J. Phys.} \textbf{31} (1963) 6--15} \expandafter\def\csname PFref@Wilce2019\endcsname{A.~Wilce, Conjugates, filters and quantum mechanics, \emph{Quantum} \textbf{3} (2019) 158. \href{https://doi.org/10.22331/q-2019-07-08-158}{doi:10.22331/q-2019-07-08-158}. \href{https://arxiv.org/abs/1206.2897}{arXiv:1206.2897}} \expandafter\def\csname PFref@Wolf1963elliptic\endcsname{J.~A. Wolf, Elliptic spaces in Grassmann manifolds, \emph{Illinois J. Math.} \textbf{7} (1963) 447--462. \href{https://doi.org/10.1215/ijm/1255644952}{doi:10.1215/i{}jm/1255644952}. Correction: D.~B. Shapiro and J.~A. Wolf, On isoclinic closure: correction to ``Elliptic spaces in Grassmann manifolds'', \emph{Illinois J. Math.} \textbf{19} (1975) 316--317. \href{https://doi.org/10.1215/ijm/1256050819}{doi:10.1215/i{}jm/1256050819}} \expandafter\def\csname PFref@Wolf1963geodesic\endcsname{J.~A. Wolf, Geodesic spheres in Grassmann manifolds, \emph{Illinois J. Math.} \textbf{7} (1963) 425--446. \href{https://doi.org/10.1215/ijm/1255644951}{doi:10.1215/i{}jm/1255644951}} \expandafter\def\csname PFref@Wong1961\endcsname{Y.-C. Wong, Isoclinic $n$-planes in Euclidean $2n$-space, Clifford parallels in elliptic $(2n-1)$-space, and the Hurwitz matrix equations, \emph{Mem. Amer. Math. Soc.} \textbf{41} (1961). \href{https://doi.org/10.1090/memo/0041}{doi:10.1090/memo/0041}} \expandafter\def\csname PFref@ZyczkowskiSommers2001\endcsname{K.~\.Zyczkowski and H.-J. Sommers, Induced measures in the space of mixed quantum states, \emph{J. Phys. A} \textbf{34} (2001) 7111--7125. \href{https://doi.org/10.1088/0305-4470/34/35/335}{doi:10.1088/0305-4470/34/35/335}. \href{https://arxiv.org/abs/quant-ph/0012101}{arXiv:quant-ph/0012101}} \expandafter\def\csname PFref@Bargmann1964\endcsname{V.~Bargmann, Note on Wigner's theorem on symmetry operations, \emph{J. Math. Phys.} \textbf{5} (1964) 862--868. \href{https://doi.org/10.1063/1.1704188}{doi:10.1063/1.1704188}} \expandafter\def\csname PFref@LeggettGarg1985\endcsname{A.~J. Leggett and A.~Garg, Quantum mechanics versus macroscopic realism: is the flux there when nobody looks?, \emph{Phys. Rev. Lett.} \textbf{54} (1985) 857--860. \href{https://doi.org/10.1103/PhysRevLett.54.857}{doi:10.1103/PhysRevLett.54.857}} \expandafter\def\csname PFref@Fritz2010\endcsname{T.~Fritz, Quantum correlations in the temporal Clauser--Horne--Shimony--Holt (CHSH) scenario, \emph{New J. Phys.} \textbf{12} (2010) 083055. \href{https://doi.org/10.1088/1367-2630/12/8/083055}{doi:10.1088/1367-2630/12/8/083055}. \href{https://arxiv.org/abs/1005.3421}{arXiv:1005.3421}} \expandafter\def\csname PFref@Spekkens2005\endcsname{R.~W. Spekkens, Contextuality for preparations, transformations, and unsharp measurements, \emph{Phys. Rev. A} \textbf{71} (2005) 052108. \href{https://doi.org/10.1103/PhysRevA.71.052108}{doi:10.1103/PhysRevA.71.052108}. \href{https://arxiv.org/abs/quant-ph/0406166}{arXiv:quant-ph/0406166}} \expandafter\def\csname PFref@SchmidEtAl2021\endcsname{D.~Schmid, J.~H. Selby, E.~Wolfe, R.~Kunjwal and R.~W. Spekkens, Characterization of noncontextuality in the framework of generalized probabilistic theories, \emph{PRX Quantum} \textbf{2} (2021) 010331. \href{https://doi.org/10.1103/PRXQuantum.2.010331}{doi:10.1103/PRXQuantum.2.010331}. \href{https://arxiv.org/abs/1911.10386}{arXiv:1911.10386}} \expandafter\def\csname PFref@Hardy2001\endcsname{L.~Hardy, Quantum theory from five reasonable axioms, \href{https://arxiv.org/abs/quant-ph/0101012}{arXiv:quant-ph/0101012} (2001)} \expandafter\def\csname PFref@MasanesMuller2011\endcsname{Ll.~Masanes and M.~P. M\"uller, A derivation of quantum theory from physical requirements, \emph{New J. Phys.} \textbf{13} (2011) 063001. \href{https://arxiv.org/abs/1004.1483}{arXiv:1004.1483}} \expandafter\def\csname PFref@CDP2011\endcsname{G.~Chiribella, G.~M. D'Ariano and P.~Perinotti, Informational derivation of quantum theory, \emph{Phys. Rev. A} \textbf{84} (2011) 012311. \href{https://doi.org/10.1103/PhysRevA.84.012311}{doi:10.1103/PhysRevA.84.012311}. \href{https://arxiv.org/abs/1011.6451}{arXiv:1011.6451}} \expandafter\def\csname PFref@FJSS1962\endcsname{D.~Finkelstein, J.~M. Jauch, S.~Schiminovich and D.~Speiser, Foundations of quaternion quantum mechanics, \emph{J. Math. Phys.} \textbf{3} (1962) 207--220. \href{https://doi.org/10.1063/1.1703794}{doi:10.1063/1.1703794}} \expandafter\def\csname PFref@Adler1995\endcsname{S.~L. Adler, \emph{Quaternionic Quantum Mechanics and Quantum Fields}, International Series of Monographs on Physics \textbf{88}, Oxford University Press, New York, 1995} \expandafter\def\csname PFref@Schwarz1988\endcsname{G.~W. Schwarz, Invariant theory of $G_2$ and $\mathrm{Spin}_7$, \emph{Comment. Math. Helv.} \textbf{63} (1988) 624--663. \href{https://doi.org/10.1007/BF02566782}{doi:10.1007/BF02566782}} \expandafter\def\csname PFref@Freudenthal1954\endcsname{H.~Freudenthal, Beziehungen der $E_7$ und $E_8$ zur Oktavenebene~I, \emph{Indag. Math.} \textbf{16} (1954) 218--230 (= Proc. Kon. Ned. Akad. Wet. A \textbf{57}). \href{https://doi.org/10.1016/S1385-7258(54)50032-6}{doi:10.1016/S1385-7258(54)50032-6}} \expandafter\def\csname PFref@JacobsonJacobson1949\endcsname{F.~D. Jacobson and N.~Jacobson, Classification and representation of semi-simple Jordan algebras, \emph{Trans. Amer. Math. Soc.} \textbf{65} (1949) 141--169. \href{https://doi.org/10.1090/S0002-9947-1949-0029367-8}{doi:10.1090/S0002-9947-1949-0029367-8}} \expandafter\def\csname PFref@Yokota2009\endcsname{I.~Yokota, Exceptional Lie groups, \href{https://arxiv.org/abs/0902.0431}{arXiv:0902.0431} (2009)} \expandafter\def\csname PFref@Fliess1974\endcsname{M.~Fliess, Matrices de Hankel, \emph{J. Math. Pures Appl.} (9) \textbf{53} (1974) 197--222} \expandafter\def\csname PFref@Schutzenberger1961\endcsname{M.-P. Sch\"utzenberger, On the definition of a family of automata, \emph{Inf. Control} \textbf{4} (1961) 245--270. \href{https://doi.org/10.1016/S0019-9958(61)80020-X}{doi:10.1016/S0019-9958(61)80020-X}} \expandafter\def\csname PFref@Lean4\endcsname{L.~de Moura and S.~Ullrich, The Lean~4 theorem prover and programming language, in \emph{Automated Deduction -- CADE~28}, Lecture Notes in Comput. Sci. \textbf{12699}, Springer, Cham, 2021, 625--635. \href{https://doi.org/10.1007/978-3-030-79876-5_37}{doi:10.1007/978-3-030-79876-5\_37}} \expandafter\def\csname PFref@Mathlib2020\endcsname{The mathlib Community, The Lean mathematical library, in \emph{Proc. 9th ACM SIGPLAN Int. Conf. on Certified Programs and Proofs (CPP 2020)}, ACM, 2020, 367--381. \href{https://doi.org/10.1145/3372885.3373824}{doi:10.1145/3372885.3373824}. \href{https://arxiv.org/abs/1910.09336}{arXiv:1910.09336}} \expandafter\def\csname PFshort@AsoreyScolariciSolombrino2007\endcsname{Asorey, Scolarici and Solombrino, \emph{Complex projections of quaternionic maps}} \expandafter\def\csname PFshort@Catto2003\endcsname{Catto, \emph{Exceptional projective geometries}} \expandafter\def\csname PFshort@MasilloScolariciSozzo2009\endcsname{Masillo, Scolarici and Sozzo, \emph{Proper versus improper mixtures}} \expandafter\def\csname PFshort@AbramskyBrandenburger2011\endcsname{Abramsky and Brandenburger, \emph{The sheaf-theoretic structure}} \expandafter\def\csname PFshort@Albert1934\endcsname{Albert, ``On a certain algebra of quantum mechanics''} \expandafter\def\csname PFshort@AlbertPaige1959\endcsname{Albert and Paige, ``On a homomorphism property of certain Jordan algebras''} \expandafter\def\csname PFshort@AlfsenShultz2003\endcsname{Alfsen and Shultz, \emph{Geometry of State Spaces}} \expandafter\def\csname PFshort@AlfsenShultzStormer1978\endcsname{Alfsen, Shultz and St{\o}rmer, \emph{A Gelfand--Neumark theorem for Jordan algebras}} \expandafter\def\csname PFshort@Araki1980b\endcsname{Araki, \emph{On a characterization of the state space}} \expandafter\def\csname PFshort@Baez2002\endcsname{Baez, ``The octonions''} \expandafter\def\csname PFshort@Baez2012\endcsname{Baez, ``Division algebras and quantum theory''} \expandafter\def\csname PFshort@BaezBokorBoyle2026\endcsname{Baez, Bokor and Boyle, \emph{Jordan pair quantum theory and the standard model}} \expandafter\def\csname PFshort@BaezSchwahn2026\endcsname{Baez and Schwahn, \emph{The Standard Model gauge group}} \expandafter\def\csname PFshort@BarnumHilgert2020\endcsname{Barnum and Hilgert, ``Spectral properties of convex bodies''} \expandafter\def\csname PFshort@BarnumWilce2014\endcsname{Barnum and Wilce, \emph{Local tomography}} \expandafter\def\csname PFshort@BGW2015\endcsname{Barnum, Graydon and Wilce, \emph{Some nearly quantum theories}} \expandafter\def\csname PFshort@BGW2020\endcsname{Barnum, Graydon and Wilce, \emph{Composites and categories}} \expandafter\def\csname PFshort@BMU2014\endcsname{Barnum, M\"uller and Ududec, \emph{Higher-order interference and single-system postulates}} \expandafter\def\csname PFshort@Boyle2020\endcsname{Boyle, \emph{The standard model, the exceptional Jordan algebra, and triality}} \expandafter\def\csname PFshort@BoyleFarnsworth2020\endcsname{Boyle and Farnsworth, \emph{The standard model, the Pati--Salam model, and `Jordan geometry'}} \expandafter\def\csname PFshort@Brehm1990\endcsname{Brehm, \emph{The shape invariant of triangles}} \expandafter\def\csname PFshort@BMKG2013\endcsname{Budroni et al., ``Bounding temporal quantum correlations''} \expandafter\def\csname PFshort@BudroniEmary2014\endcsname{Budroni and Emary, \emph{Temporal quantum correlations}} \expandafter\def\csname PFshort@BudroniFagundesKleinmann2019\endcsname{Budroni, Fagundes and Kleinmann, \emph{Memory cost of temporal correlations}} \expandafter\def\csname PFshort@BuschGrabowskiLahti1995\endcsname{Busch, Grabowski and Lahti, \emph{Repeatable measurements in quantum theory}} \expandafter\def\csname PFshort@BuschSingh1998\endcsname{Busch and Singh, ``L\"uders theorem for unsharp quantum measurements''} \expandafter\def\csname PFshort@ChiribellaYuan2014\endcsname{Chiribella and Yuan, \emph{Measurement sharpness}} \expandafter\def\csname PFshort@ChiribellaYuan2018\endcsname{Chiribella and Yuan, \emph{Bridging the gap}} \expandafter\def\csname PFshort@Cho2019\endcsname{Cho, \emph{Effectuses in Categorical Quantum Foundations}} \expandafter\def\csname PFshort@ClementeKofler2016\endcsname{Clemente and Kofler, ``No fine theorem for macrorealism: limitations of the Leggett--Garg inequality''} \expandafter\def\csname PFshort@DaviesLewis1970\endcsname{Davies and Lewis, ``An operational approach to quantum probability''} \expandafter\def\csname PFshort@DuboisViolette2016\endcsname{Dubois-Violette, ``Exceptional quantum geometry and particle physics''} \expandafter\def\csname PFshort@DuboisVioletteTodorov2018\endcsname{Dubois-Violette and Todorov, ``Exceptional quantum geometry and particle physics~II''} \expandafter\def\csname PFshort@Ehrlich2026\endcsname{Ehrlich, \emph{\texttt{sequential-products-lean}}} \expandafter\def\csname PFshort@Ehrlich2026ca\endcsname{Ehrlich, \emph{\texttt{composition-algebras}}} \expandafter\def\csname PFshort@Ehrlich2026ej\endcsname{Ehrlich, \emph{\texttt{euclidean-jordan}}} \expandafter\def\csname PFshort@ElduqueIltyakov1999\endcsname{Elduque and Iltyakov, \emph{Polynomial invariants of exceptional simple algebraic groups}} \expandafter\def\csname PFshort@FLW2024\endcsname{Fanizza, Lumbreras and Winter, \emph{Quantum theory in finite dimension}} \expandafter\def\csname PFshort@FNW1992\endcsname{Fannes, Nachtergaele and Werner, \emph{Finitely correlated states}} \expandafter\def\csname PFshort@FarautKoranyi1994\endcsname{Faraut and Kor\'anyi, \emph{Analysis on Symmetric Cones}} \expandafter\def\csname PFshort@FernandesWagnerNovoGalvao2024\endcsname{Fernandes et al., \emph{Unitary-invariant witnesses of quantum imaginarity}} \expandafter\def\csname PFshort@GarnerMuller2020\endcsname{Garner and M\"uller, \emph{Characterization of the probabilistic models}} \expandafter\def\csname PFshort@Glennie1966\endcsname{Glennie, ``Some identities valid in special Jordan algebras''} \expandafter\def\csname PFshort@GMS2025\endcsname{Gorshkov, Mamontov and Staroletov, \emph{Jordan algebras that are factors of Matsuo algebras}} \expandafter\def\csname PFshort@GorshkovStaroletov2020\endcsname{Gorshkov and Staroletov, \emph{Primitive 3-generated axial algebras}} \expandafter\def\csname PFshort@GudderGreechie2002\endcsname{Gudder and Greechie, ``Sequential products on effect algebras''} \expandafter\def\csname PFshort@GPR1978\endcsname{G\"unaydin, Piron and Ruegg, \emph{Moufang plane and octonionic quantum mechanics}} \expandafter\def\csname PFshort@GunaydinGursey1973\endcsname{G\"unaydin and G\"ursey, ``Quark structure and octonions''} \expandafter\def\csname PFshort@HarveyLawson1982\endcsname{Harvey and Lawson, ``Calibrated geometries''} \expandafter\def\csname PFshort@HeinosaariReitznerStano2008\endcsname{Heinosaari, Reitzner and Stano, \emph{Notes on joint measurability}} \expandafter\def\csname PFshort@HSGB2018\endcsname{Hoffmann et al., \emph{Structure of temporal correlations of a qubit}} \expandafter\def\csname PFshort@Hoggar1977\endcsname{Hoggar, ``New sets of equi-isoclinic $n$-planes from old''} \expandafter\def\csname PFshort@Hsiang1989\endcsname{Hsiang, \emph{Laws of trigonometries}} \expandafter\def\csname PFshort@Iltyakov1997\endcsname{Iltyakov, \emph{Polynomial invariants of group $F_4$}} \expandafter\def\csname PFshort@JordanvNWigner1934\endcsname{Jordan, von Neumann and Wigner, \emph{On an algebraic generalization}} \expandafter\def\csname PFshort@KhandelwalTavakoli2025\endcsname{Khandelwal and Tavakoli, \emph{Simulating quantum instruments}} \expandafter\def\csname PFshort@Kleinmann2014\endcsname{Kleinmann, \emph{Sequences of projective measurements}} \expandafter\def\csname PFshort@Kollross\endcsname{Kollross, \emph{Octonions, triality, the exceptional Lie algebra $F_4$}} \expandafter\def\csname PFshort@Krasnov2019\endcsname{Krasnov, ``SO(9) characterization of the Standard Model gauge group''} \expandafter\def\csname PFshort@Luders1951\endcsname{L\"uders, ``\"Uber die Zustands\"anderung durch den Me{\ss}proze{\ss}''} \expandafter\def\csname PFshort@Physlib\endcsname{Diem, \emph{\texttt{physlib}}} \expandafter\def\csname PFshort@MacGregorPhi2026\endcsname{MacGregor, ``Finite observations and recognition of the compact Albert algebra''} \expandafter\def\csname PFshort@McCrimmon2004\endcsname{McCrimmon, \emph{A Taste of Jordan Algebras}} \expandafter\def\csname PFshort@McCrimmonIdentities2004\endcsname{McCrimmon, ``The role of identities in Jordan algebras''} \expandafter\def\csname PFshort@MonrasWinter2016\endcsname{Monr\`as and Winter, \emph{Quantum learning of classical stochastic processes}} \expandafter\def\csname PFshort@MullerGarner2023\endcsname{M\"uller and Garner, \emph{Testing quantum theory}} \expandafter\def\csname PFshort@NPA2008\endcsname{Navascu\'es, Pironio and Ac\'in, \emph{A convergent hierarchy}} \expandafter\def\csname PFshort@Niestegge2008\endcsname{Niestegge, \emph{Quantum measurement in order-unit spaces}} \expandafter\def\csname PFshort@Niestegge2012\endcsname{Niestegge, \emph{Conditional probability, three-slit experiments}} \expandafter\def\csname PFshort@Niestegge2020a\endcsname{Niestegge, \emph{A simple and quantum-mechanically motivated characterization}} \expandafter\def\csname PFshort@Niestegge2020b\endcsname{Niestegge, \emph{Quantum probability's algebraic origin}} \expandafter\def\csname PFshort@Niestegge2026a\endcsname{Niestegge, \emph{Homogeneity of the quantum transition probability}} \expandafter\def\csname PFshort@Nasmith2022\endcsname{Nasmith, ``Octonions and the two strictly projective tight 5-designs''} \expandafter\def\csname PFshort@Nasmith2022integers\endcsname{Nasmith, \emph{Octonion integers and tight 5-designs}} \expandafter\def\csname PFshort@Hoehn2026lattice\endcsname{H\"ohn, \emph{The unique extremal threemodular lattice of rank 26}} \expandafter\def\csname PFshort@Hoehn2026albert\endcsname{H\"ohn, \emph{Definite integral Albert algebras}} \expandafter\def\csname PFshort@Niestegge2026b\endcsname{Niestegge, \emph{The bare necessities}} \expandafter\def\csname PFshort@OrtegaSantander2002\endcsname{Ortega and Santander, \emph{Trigonometry of `complex Hermitian' type}} \expandafter\def\csname PFshort@OszmaniecBrodGalvao2021\endcsname{Oszmaniec, Brod and Galv\~ao, \emph{Measuring relational information}} \expandafter\def\csname PFshort@Ozawa1984\endcsname{Ozawa, ``Quantum measuring processes of continuous observables''} \expandafter\def\csname PFshort@Peres1979\endcsname{Peres, ``Proposed test for complex versus quaternion quantum theory''} \expandafter\def\csname PFshort@MMG2009\endcsname{McKague, Mosca and Gisin, ``Simulating quantum systems using real Hilbert spaces''} \expandafter\def\csname PFshort@Renou2021\endcsname{Renou et al., ``Quantum theory based on real numbers can be experimentally falsified''} \expandafter\def\csname PFshort@PinedaRua2026\endcsname{Pineda Montoya and Rua Munoz, \emph{Operational foundations}} \expandafter\def\csname PFshort@PNA2010\endcsname{Pironio, Navascu\'es and Ac\'in, \emph{Convergent relaxations}} \expandafter\def\csname PFshort@PopovaSamoilenko2006\endcsname{Popova and Samoilenko, \emph{Configurations of subspaces with fixed angles}} \expandafter\def\csname PFshort@Racine1977\endcsname{Racine, ``Maximal subalgebras of exceptional Jordan algebras''} \expandafter\def\csname PFshort@RoelandsTiersma2026\endcsname{Roelands and Tiersma, ``On the Shirshov--Cohn theorem for JB-algebras''} \expandafter\def\csname PFshort@SahaSantosAugusiak2020\endcsname{Saha, Santos and Augusiak, \emph{Sum-of-squares decompositions}} \expandafter\def\csname PFshort@SalamonWalpuski2017\endcsname{Salamon and Walpuski, ``Notes on the octonions''} \expandafter\def\csname PFshort@SamoilenkoYershova2006\endcsname{Samoilenko and Yershova, \emph{Equiangular configurations of subspaces}} \expandafter\def\csname PFshort@SlomczynskiSzczepanek2018\endcsname{S{\l}omczy\'nski and Szczepanek, \emph{Quantum dynamical entropy}} \expandafter\def\csname PFshort@Soares2026\endcsname{Soares, \emph{The Albert algebra in Lean~4}} \expandafter\def\csname PFshort@Sorkin1994\endcsname{Sorkin, ``Quantum mechanics as quantum measure theory''} \expandafter\def\csname PFshort@SpringerVeldkamp2000\endcsname{Springer and Veldkamp, \emph{Octonions, Jordan Algebras and Exceptional Groups}} \expandafter\def\csname PFshort@SupicBowles2020\endcsname{\v{S}upi\'c and Bowles, ``Self-testing of quantum systems: a review''} \expandafter\def\csname PFshort@UdudecBarnumEmerson2011\endcsname{Ududec, Barnum and Emerson, \emph{Three slit experiments}} \expandafter\def\csname PFshort@vandeWetering\endcsname{van de Wetering, ``Three characterisations of the sequential product''} \expandafter\def\csname PFshort@vandeWetering2019a\endcsname{van de Wetering, ``Sequential product spaces are Jordan algebras''} \expandafter\def\csname PFshort@VitaglianoBudroni2022\endcsname{Vitagliano and Budroni, ``Leggett--Garg macrorealism and temporal correlations''} \expandafter\def\csname PFshort@Vorobev1962\endcsname{Vorob'ev, ``Consistent families of measures and their extensions''} \expandafter\def\csname PFshort@WWvdW2022\endcsname{Westerbaan and van de Wetering, ``A computer scientist's reconstruction of quantum theory''} \expandafter\def\csname PFshort@WWvdW2020\endcsname{Westerbaan, Westerbaan and van de Wetering, ``The three types of normal sequential effect algebras''} \expandafter\def\csname PFshort@Wigner1963\endcsname{Wigner, ``The problem of measurement''} \expandafter\def\csname PFshort@Wilce2019\endcsname{Wilce, ``Conjugates, filters and quantum mechanics''} \expandafter\def\csname PFshort@Wolf1963elliptic\endcsname{Wolf, ``Elliptic spaces in Grassmann manifolds''} \expandafter\def\csname PFshort@Wolf1963geodesic\endcsname{Wolf, ``Geodesic spheres in Grassmann manifolds''} \expandafter\def\csname PFshort@Wong1961\endcsname{Wong, \emph{Isoclinic $n$-planes}} \expandafter\def\csname PFshort@ZyczkowskiSommers2001\endcsname{\.Zyczkowski and Sommers, \emph{Induced measures}} \expandafter\def\csname PFshort@Bargmann1964\endcsname{Bargmann, ``Note on Wigner's theorem''} \expandafter\def\csname PFshort@LeggettGarg1985\endcsname{Leggett and Garg, ``Quantum mechanics versus macroscopic realism''} \expandafter\def\csname PFshort@Fritz2010\endcsname{Fritz, ``Quantum correlations in the temporal CHSH scenario''} \expandafter\def\csname PFshort@Spekkens2005\endcsname{Spekkens, ``Contextuality for preparations''} \expandafter\def\csname PFshort@SchmidEtAl2021\endcsname{Schmid et al., ``Characterization of noncontextuality''} \expandafter\def\csname PFshort@Hardy2001\endcsname{Hardy, ``Quantum theory from five reasonable axioms''} \expandafter\def\csname PFshort@MasanesMuller2011\endcsname{Masanes and M\"uller, ``A derivation of quantum theory''} \expandafter\def\csname PFshort@CDP2011\endcsname{Chiribella, D'Ariano and Perinotti, ``Informational derivation''} \expandafter\def\csname PFshort@FJSS1962\endcsname{Finkelstein et al., ``Foundations of quaternion quantum mechanics''} \expandafter\def\csname PFshort@Adler1995\endcsname{Adler, \emph{Quaternionic Quantum Mechanics}} \expandafter\def\csname PFshort@Schwarz1988\endcsname{Schwarz, ``Invariant theory of $G_2$ and $\mathrm{Spin}_7$''} \expandafter\def\csname PFshort@Freudenthal1954\endcsname{Freudenthal, ``Beziehungen der $E_7$ und $E_8$ zur Oktavenebene~I''} \expandafter\def\csname PFshort@JacobsonJacobson1949\endcsname{Jacobson and Jacobson, ``Classification and representation''} \expandafter\def\csname PFshort@Yokota2009\endcsname{Yokota, \emph{Exceptional Lie groups}} \expandafter\def\csname PFshort@Fliess1974\endcsname{Fliess, ``Matrices de Hankel''} \expandafter\def\csname PFshort@Schutzenberger1961\endcsname{Sch\"utzenberger, ``On the definition of a family of automata''} \expandafter\def\csname PFshort@Lean4\endcsname{de Moura and Ullrich, ``The Lean~4 theorem prover''} \expandafter\def\csname PFshort@Mathlib2020\endcsname{The mathlib Community, ``The Lean mathematical library''} \makeatletter \let\PFfootnote\footnote \newcommand{\PFlastcite}{} \pretocmd{\footnote}{\gdef\PFlastcite{}}{}{} \newcounter{PFnoterefcount} \newcommand{\PFciteone}[1]{\ifcsdef{PFref@#1}{}{\PackageWarning{PFcite}{No reference text for #1}}\PFfirstfalse \ifcsdef{PFseen@#1}{\ifPFintable\edef\PFtmp{\thetable}\ifcsequal{PFfirsttab@#1}{PFtmp}{\PFfirsttrue}{}\fi}{\PFfirsttrue}\ifPFfirst \Hy@raisedlink{\hypertarget{PFcite.#1}{}}\csname PFref@#1\endcsname \csxdef{PFptr@#1}{\PFhere}\csgdef{PFseen@#1}{1}\ifPFintable\csxdef{PFfirsttab@#1}{\thetable}\fi \else \csname PFshort@#1\endcsname, \hyperlink{PFcite.#1}{\csname PFptr@#1\endcsname}\fi} \newif\ifPFfirst\newif\ifPFintable \newcommand{\PFnoteref}[1]{\ifnum\value{PFnoterefcount}>0 ; \fi \stepcounter{PFnoterefcount}\PFciteone{#1}} \newcommand{\PFhere}{} \renewcommand{\cite}[2][]{\unskip\PFfootnote{\edef\PFhere{note~\the\value{footnote}}\def\PFcurrentcite{#2}\ifx\PFcurrentcite\PFlastcite Ibid.\else \setcounter{PFnoterefcount}{0}\forcsvlist{\PFnoteref}{#2}.\fi \if\relax\detokenize{#1}\relax\else\space #1.\fi \setcounter{PFnoterefcount}{0}\forcsvlist{\PFcountcite}{#2}\ifnum\value{PFnoterefcount}=1 \gdef\PFlastcite{#2}\else\gdef\PFlastcite{}\fi}} \newcommand{\PFcountcite}[1]{\stepcounter{PFnoterefcount}} \newcommand{\PFtablecite}[2][]{\edef\PFhere{Table~\thetable}\setcounter{PFnoterefcount}{0}[\forcsvlist{\PFnoteref}{#2}\if\relax\detokenize{#1}\relax\else, #1\fi]} \AtBeginEnvironment{table}{\let\cite\PFtablecite\PFintabletrue\gdef\PFlastcite{}} \AtEndEnvironment{table}{\gdef\PFlastcite{}} \makeatother \begin{document} \newgeometry{hmargin={1.55in,1.55in},top=1.1in,bottom=1.0in} \pagenumbering{roman} \thispagestyle{empty} \begingroup \raggedright \setlength{\parindent}{0pt} \setlength{\parskip}{0.55\baselineskip} {\sffamily\footnotesize AI proof note\hfill 18 November 2026}\par \vspace{0.9in} {\Large On sequential ideal measurements\\[0.25ex] in the exceptional Jordan algebra\par} \vspace{0.12in} {\large Robert R. MacGregor\par} \vspace{0.5in} \coverhead{Abstract} Real, complex and quaternionic quantum theory have one exceptional relative, the Albert algebra $H_3(\mathbb O)$, which single-system reconstructions leave standing. Sequences of measurements on a single system can still tell it apart, provided the measurements are ideal. A yes--no test is ideal if it disturbs the state no more than its answer requires, as under the projection postulate. Ideal tests of finitely many pure states of $H_3(\mathbb O)$ are then matched by ideal measurements on a real, complex or quaternionic Hilbert space of any dimension exactly when those states fail to generate the algebra. When they generate it, an ideal Hilbert-space model matching every sequence from even one starting state would yield a faithful representation of the algebra and its Jordan product by operators, which Albert ruled out in 1934. Three states never generate it; whether four do is decided by one number, proportional to a squared octonion associator. Even when four do, the difference is slow to show: from the maximally mixed state, a random choice between two quaternionic qutrits matches every sequence of up to six outcomes, and only at seven is every Hilbert-space model excluded. If the tests need not be ideal, then for any finite set of tests a finite-dimensional quantum device with sharp answers matches every sequence, of any length. \vspace{0.3in} \coverhead{AI Statement} All text, mathematics and proofs herein, and all Lean~4 code apart from the adapted files noted in Appendix~\ref{app:lean-plan}, were produced by Claude Fable~5.2, Claude Opus~5.5, GPT-6.2 Astra, GPT-6.2 Sol, Gemini Carbon and Gemini Hydrogen, and the main theorems on the Albert algebra are machine-checked in Lean. \emph{Except for this statement, all content herein was generated by AI and has not been understood without AI interpretation.} \vfill {\sffamily\footnotesize Text CC BY 4.0\quad Lean code Apache-2.0\\ Lean repository: \leanrepo\\ ORCID: \href{https://orcid.org/0009-0003-1734-8054}{0009-0003-1734-8054}\par} \endgroup \clearpage \restoregeometry \pagenumbering{arabic} \pagenumbering{roman}\setcounter{page}{2} {\small\tableofcontents} \clearpage \pagenumbering{arabic} \section{Introduction}\label{sec:intro} \leannoteexplain{Margin notes name the Lean declaration that proves the adjacent statement, with a remark where the formal route or scope differs; algebraic models are defined in Appendix~\ref{app:lean-plan}.}Exceptional algebraic structure can stay invisible in restricted descriptions and in short sequences of measurements, and become detectable once one description has to stay consistent under repeated ideal measurement. A precise form of this holds for the Albert algebra $\J=\HO$ of Hermitian $3\times3$ matrices over the octonions $\Oct$, with the symmetrised product $x\jp y=\frac12(xy+yx)$. It is the one exceptional system that single-system reconstructions of quantum theory leave standing (Section~\ref{sec:intro:exception}). The statement has two halves. Take four ideal yes--no tests on $\J$ that generate the algebra, meaning that sums and symmetrised products of the four tested elements and the unit $e$ reach all of $\J$, and start from the maximally mixed state $e/3$, the analogue of the density matrix $I/3$. Through six tests in sequence, the Albert statistics are those of a random choice between two quaternionic quantum systems (Proposition~\ref{prop:cer:universalsix}). At seven they are not. A round trip, four tests run forward and then back to the first, has seven steps and measures a square: a random choice raises it by the variance between the two systems, while the Albert value is the square of their mean (paragraph ``Why the mixture fails at seven'' in Section~\ref{sec:cer:horizon}). No Hilbert-space model with ideal measurements, of any dimension, matches every sequence of length seven (Theorem~\ref{thm:cer:uniform}). Behind both halves stands a criterion. A Hilbert-space model is asked only to reproduce probabilities, yet reproducing all of them forces a Jordan homomorphism, a linear map preserving the product $\jp$, from the algebra the tests generate into operators (Theorem~\ref{thm:bridgeJ}). Ideality, the requirement that each test disturb the state only as much as its answer requires (for sharp tests the L\"uders rule, Theorem~\ref{thm:ideal}), is an assumption of the comparison and cannot be certified from the statistics. A quantum device whose tests disturb the state after answering reproduces every word probability of the Albert tests, however long the words (Theorem~\ref{thm:cp69}, Corollary~\ref{cor:ins:sharpalbert}), so no table of word probabilities shows that its tests were ideal. The results compare two model classes under the same postulate, ideal tests on the Albert algebra and ideal tests on Hilbert spaces, and Section~\ref{sec:ins} shows that the postulate does real work: without it the exceptional statistics have an ordinary quantum implementation. Nothing here bears on whether nature is quantum, and no complexity claim is made: $\J$ is a $27$-dimensional real vector space on which the ideal updates act as linear maps. \subsection{Four buttons and a light}\label{sec:intro:box} Imagine a sealed box with four buttons and a light that answers yes or no. Each button is an ideal test of an \emph{atom} of $\J$. An atom plays the part that a rank-one projector plays in quantum theory: a yes--no question that cannot be split into finer ones, and equally the pure state in which its answer is certainly yes. The four atoms are chosen to generate $\J$, and the box starts in the maximally mixed state $e/3$. Someone who has only the buttons can estimate the probability of every sequence of answers. What ``ideal'' means is easiest to see on an ordinary quantum particle with three paths. Leave path three empty and let paths one and two interfere. Two detectors can both answer ``not on path three'': one leaves the other two paths alone, the other quietly records which of them was taken. Only the first preserves the interference. An ideal test changes the state only as much as its answer requires, and in quantum theory it is the L\"uders rule $\rho\mapsto P\rho P$ (Theorem~\ref{thm:ideal}). An ordinary quantum device that may disturb the state after answering reproduces every sequence of the four Albert tests, however long, and its tests are even sharp and repeatable: each answer is given by a projection, and an immediate repetition confirms it (Corollary~\ref{cor:ins:sharpalbert}). Require each of its tests to make only the ideal change, and some prediction fails by the seventh press (Theorem~\ref{thm:cer:uniform}). The schedule matters too. Choose any seven-press schedule, a fixed sequence of seven of the four tests; an ordinary quantum device whose tests are ideal can be built to agree with the box on that schedule and on every run of six presses or fewer, and some other seven-press schedule must then give a different probability (Proposition~\ref{prop:cer:schedule}(a) and Theorem~\ref{thm:cer:uniform}). \subsection{A warm-up with a qubit and a qutrit}\label{sec:intro:warmup} Take a qubit in the state $\rho$ and two yes--no measurements given by rank-one projections $P_1$, $P_2$ on $\C^2$, with transition probability $o=\Tr(P_1P_2)$. By the L\"uders rule~\cite{Luders1951} the outcome ``yes'' leaves $P_1\rho P_1$ and has probability $\Tr(\rho P_1)$. Since $P_1P_2P_1=o\,P_1$, a ``yes'' on $P_2$ followed by a ``yes'' on $P_1$ multiplies this probability by $o^2$: \[ \Pr(\text{yes}_1,\text{yes}_2,\text{yes}_1)=\Tr(\rho P_1)\,o^2 . \] A ``yes'' on a rank-one projection forgets everything about the past except a number. A ``no'' does not. On a qutrit with rank-one $P_1$, the outcome ``no'' updates by the rank-two projection $1-P_1$ and leaves a two-dimensional piece of the state coherent. Three measurements on a qutrit, with ``no'' outcomes allowed, see the phase of the triangle formed by their unit vectors $a_1,a_2,a_3$. Writing $\Pr(i,\bar\jmath,k)$ for the probability of ``yes'' on $i$, ``no'' on $j$, ``yes'' on $k$, from the maximally mixed state, and $\Pr(i,k)$ for ``yes'' on $i$ then ``yes'' on $k$, \[ \operatorname{Re}\langle a_i|a_j\rangle\langle a_j|a_k\rangle\langle a_k|a_i\rangle =\tfrac32\bigl[\Pr(i,k)+\Pr(i,j,k)-\Pr(i,\bar\jmath,k)\bigr], \] the real part of the Bargmann invariant of the three states~\cite{Bargmann1964}. The reason is that the anticommutator is a signed sum of the identity and the two L\"uders updates, $\tfrac12(PX+XP)=\tfrac12\bigl(X+PXP-(1-P)X(1-P)\bigr)$. In this sense sequences of ideal measurements measure Jordan products. A Jordan algebra is a real vector space with a commutative product $\jp$ of this kind (Section~\ref{sec:set:eja}), and in it the identity reads $p\jp x=\tfrac12(x+U_px-U_{e-p}x)$, where $U_p$ and $U_{e-p}$ are the ideal updates (Section~\ref{sec:set}), and the right side makes sense for $\J$. To fit such statistics in a Hilbert space, fix finitely many measurements and ask for a Hilbert space, one projection $P_i$ per measurement and a state $\varphi$ such that every \emph{word}, a finite sequence of outcomes in chronological order, has the same probability as in the system: \[ \varphi\bigl(Q_1Q_2\cdots Q_m\cdots Q_2Q_1\bigr)=\Pr(q_1,\dots,q_m),\qquad Q_t\in\{P_i,1-P_i\}. \] These are the sequential projective representations of Budroni, Moroder, Kleinmann and G\"uhne~\cite{BMKG2013}, here allowed to be real, complex or quaternionic and to use any state (model class~(I), Definition~\ref{def:idealmodel}). The rank of $P_i$ must be free: a comparison restricted to rank-one projections would declare the quaternionic qutrit, with its rank-two projections on $\C^6$, non-quantum. Can ideal yes--no measurements on $\J$ be modelled in this way, in any dimension and with projections of any rank? For four pure states that generate $\J$, the answer depends on how the instruments are modelled: \begin{center}\small \begin{tabularx}{0.92\linewidth}{@{}>{\raggedright\arraybackslash}Xl@{}} \toprule Instruments allowed in the quantum model & Reproduces all word probabilities?\\ \midrule One projection per test, L\"uders update on both outcomes (class (I)) & no, in any dimension (Thm.~\ref{thm:bridgeJ})\\ One projection per test, arbitrary post-measurement channels & yes, on $\C^{138}$; on $\C^{136}$ from $e/3$ (Cor.~\ref{cor:ins:sharpalbert})\\ Arbitrary completely positive instruments (class (II)) & yes, on $\C^{69}$; on $\C^{68}$ from $e/3$ (Thm.~\ref{thm:cp69}, Cor.~\ref{cor:ins:nohalf})\\ \bottomrule \end{tabularx} \end{center} \noindent The exclusion is relative to class (I). A model with arbitrary instruments in infinite dimension reproduces any process (Section~\ref{sec:set:comp}). Asking less reveals more. If each yes--no test is replaced by a complete three-outcome measurement of a Jordan frame, three mutually orthogonal atoms summing to $e$ (the analogue of an orthonormal basis), the statistics become those of a classical Markov chain in the transition probabilities, and the exception disappears (Proposition~\ref{prop:markov}). \subsection{The exception that single-system axioms leave standing}\label{sec:intro:exception} Reconstructions of quantum theory from postulates about a single system narrow the possibilities to the finite-dimensional formally real (also called Euclidean) Jordan algebras, in which a sum of squares vanishes only if every term does. Jordan, von Neumann and Wigner~\cite{JordanvNWigner1934} classified them. The list consists of the real, complex and quaternionic quantum systems $H_n(\R)$, $H_n(\C)$, $H_n(\Ha)$, the spin factors, whose state spaces are balls like the Bloch ball of a qubit, and one more: the $27$-dimensional Albert algebra $\J$, also called the octonionic qutrit~\cite{Baez2002,BaezSchwahn2026}. It is \emph{exceptional}: Albert~\cite{Albert1934} showed that it is not isomorphic to a Jordan algebra of self-adjoint operators on any Hilbert space. The classical algebraic witness of this is Glennie's identity~\cite{Glennie1966}, an equation of degree eight that holds in every special Jordan algebra, that is, in every Jordan algebra of operators, and fails in $\J$. The known ways of removing $\J$ use more than one system, dynamics, or the whole effect space, and the single-system interference tests that exist do not see it (Section~\ref{sec:prior:recon}). The Albert algebra serves here as a \emph{mathematical foil}, a consistent comparison case rather than a proposed physical theory. Every table, the list of all word probabilities, of ideal tests on a Euclidean Jordan algebra is a finite convex combination of finite-dimensional quantum tables and Albert tables (Remark~\ref{rem:bri:foil}); hence a finite linear inequality valid for ideal quantum measurements either holds for ideal tests on every Euclidean Jordan algebra or is violated by an Albert table. The Albert algebra is the unique simple exceptional ingredient in this decomposition, not a unique countermodel. Up to three tests cannot separate the classes, while four can (Corollary~\ref{cor:four}). Glennie's identity separates $\J$ from the special algebras by an equation. In the word statistics the first separation is an inequality. At four atoms the element $F$ obtained from Glennie's identity has squared length proportional to the invariant $\Acal$ of result~2 below, with a factor fixed by three of the atoms (Theorem~\ref{thm:glennie}). \subsection{Results}\label{sec:intro:results} A \emph{test} is an ideal two-outcome measurement of an idempotent, an element $p$ with $p\jp p=p$ (the analogue of a projection). In a Euclidean Jordan algebra, a two-outcome instrument that leaves undisturbed every state on which its outcome is certain, and is sharp in the sense of Definition~\ref{def:ideal}, is the L\"uders instrument $(U_p,U_{e-p})$ of an idempotent; the same holds in quantum theory over $\R$, $\C$ or $\Ha$ in any dimension (Theorem~\ref{thm:ideal}). The six transition probabilities of four atoms are $o_{ij}=\tr(p_i\jp p_j)$, with $\tr$ the trace of $\J$ (Section~\ref{sec:set:eja}). \begin{enumerate}[label=\textbf{\arabic*.},leftmargin=1.6em,itemsep=4pt] \item \emph{A sequential criterion for speciality.} Tests of atoms of $\J$ have a sequential projective representation if and only if the atoms generate a proper subalgebra, and then from every state (Theorem~\ref{thm:bridgeJ}). For general Euclidean Jordan algebras the criterion is that the state gives no weight to a summand isomorphic to $\J$ in the algebra the tests generate, under a primitivity hypothesis on the tests (Theorem~\ref{thm:bridge}). An algebra is special if and only if every four of its tests are representable from every state (Corollary~\ref{cor:four}), and when the dimension is not fixed the statistics do not separate real, complex and quaternionic quantum theory (Proposition~\ref{prop:field}). \item \emph{Three tests fit a qutrit; four need not.} Any three atoms of $\J$ lie in one complex subalgebra $W\cong\HC$, so three tests are reproduced by an ordinary qutrit with rank-one projections (Proposition~\ref{prop:three}). For four atoms a single number $\Acal\ge0$, built from thirteen traces of Jordan products, vanishes if and only if the atoms lie in a quaternionic subalgebra and is positive if and only if they generate $\J$; in a normal form it is a squared octonion associator $(xy)z-x(yz)$ (Theorem~\ref{thm:fourpoint}). Seen from the subalgebra $W$ generated by three of the atoms, when it is complex (for instance when the triangle factor $\Delta_W$ of the three atoms is nonzero), $\Acal=\Delta_W\cdot N(P_Wp_4)$, the triangle factor times the determinant $N$ of the fourth atom projected onto $W$ (Proposition~\ref{prop:windowdet}). The maximum is $4/19683$, and at a maximiser each atom looks maximally mixed to the tests of the other three (Proposition~\ref{prop:max}, Corollary~\ref{cor:fp:maxmixed}). \item \emph{The tetrahedron identity.} In every sequential projective representation of four tests, the relations forced by the statistics give Clifford relations among three triangle holonomies, operators such as $P_1P_2P_3P_1$ that go round a triangle of tests (the operator form of a Bargmann phase), and associativity of operator products then forces the tetrahedron statistic $\Dtet$, a function of the probabilities of words of length at most four, to vanish (Theorem~\ref{thm:tetra}). For atoms of $\J$, $\Dtet\prod o_{ij}=\Acal$ (Proposition~\ref{prop:Dacal}). A positive $\Dtet$ alone is not a witness: for an explicit generating quadruple $\mathsf S$ (Section~\ref{sec:cer}) a representation on $\C^{12}$ reproduces all thirteen traces (Proposition~\ref{prop:Dfails}), and the companion paper gives a real representation with this property for every generating quadruple. \item \emph{Finite certificates.} For every four atoms that generate $\J$, from the maximally mixed state $e/3$, words of length at most seven separate the octonionic word probabilities from every sequential projective representation (Theorem~\ref{thm:cer:uniform}). Conversely a representation on $\C^{12}$, a mixture of two quaternionic ones, matches every word of length at most six, so the minimum separating length is $L_{\min}=7$, and no linear equality on those words separates the table (Proposition~\ref{prop:cer:universalsix}). Every sequential projective representation misses some word probability of length at most seven by an amount bounded below by a positive constant that depends on the quadruple (Theorem~\ref{thm:cer:uniform}(b)). For two explicit configurations $\mathsf M$ and $\mathsf S$ a single vector of forced negative squared norm suffices (Theorem~\ref{thm:cer:seven}). A companion paper, \emph{Finite certificates for sequential measurements in the exceptional Jordan algebra}, gives explicit values of the constant for them, together with further finite-certificate details, robustness constants and bounds for five tests. An explicit list of $2{,}041$ words of length at most eleven, the same for every generating quadruple, suffices from the maximally mixed state, and $8{,}164$ words from every state (Theorem~\ref{thm:eleven}, whose proof is machine-checked). No linear equality separates through length five for any four atoms from any state either (Remark~\ref{rem:cer:fivestep}, proved in the companion paper and in Lean); the question at greater lengths is open. \item \emph{Five tests, every four quantum.} There are five atoms of $\J$ such that every four of their tests have a sequential projective representation on $\C^6$, while the five together have none, in any dimension and from any state (Theorem~\ref{thm:pentad}). Four is the least number of tests that can show the obstruction, and five is the number that must be inspected in the worst case: a family of tests of atoms is representable as soon as every subfamily of at most five is (Proposition~\ref{prop:arity}, Corollary~\ref{cor:helly}); in the language of convexity, five is a Helly number. Words of length at most eleven exclude the five tests, by the length-seven argument for four atoms obtained from them by a reflection, carried back to the five tests; this bound is proved in the companion paper and machine-checked (Appendix~\ref{app:lean-plan}). \item \emph{Where the projection postulate enters.} With arbitrary completely positive instruments, the statistics of $n$ tests of atoms have an exact realization on $\C^{17n+1}$ (Theorem~\ref{thm:cp69}), and with sharp effects and arbitrary post-measurement channels one on $\C^{2(17n+1)}$ (Proposition~\ref{prop:ins:sharpdilation}, Corollary~\ref{cor:ins:sharpalbert}), a dimension that the companion paper lowers to $26n-7$. So the hypothesis that does the work is the L\"uders update on both outcomes, not projectivity of the effects. Conversely, with sharp effects, agreement on the words of length at most seven forces some test to disturb a state on which its answer is certain (Proposition~\ref{prop:cer:sharpdisturb}), and instruments calibrated on their own yes and no subspaces to within a tolerance set by the margin of Theorem~\ref{thm:cer:uniform}(b) (Section~\ref{sec:cer:robust}) miss the octonionic table on some word of length at most seven (Proposition~\ref{prop:cer:nearideal} and \eqref{eq:cer:nearidealuniform}). \end{enumerate} \subsection{Relation to earlier work}\label{sec:intro:known} The algebraic half of the criterion, that a JB-algebra (a complete normed generalisation of a Euclidean Jordan algebra) is special if and only if it admits no surjective Jordan homomorphism onto $\J$, is due to Alfsen, Shultz and St{\o}rmer~\cite{AlfsenShultzStormer1978} and Roelands and Tiersma~\cite[Thm.~2.11]{RoelandsTiersma2026}; it is combined here with the operational half, in which matching statistics produce a Jordan homomorphism from the generated algebra into operators (Section~\ref{sec:bri:recon}). That every triple of atoms lies in a complex subalgebra was asserted without proof by Ortega and Santander~\cite{OrtegaSantander2002}; a proof is given in Section~\ref{sec:fp:three}. The Clifford mechanism behind the tetrahedron identity is due to Wong and Wolf~\cite{Wong1961,Wolf1963geodesic,Wolf1963elliptic} for isoclinic subspaces; Section~\ref{sec:fp:tetra} adapts it to arbitrary representations. Albert's theorem~\cite{Albert1934}, that $\J$ is not a Jordan algebra of operators, enters only as the last step of the criterion (Theorem~\ref{thm:tet:albert}). The step before it assumes only that a model with ideal tests on a Hilbert space reproduces the word probabilities from one state of $\J$, in any dimension, with projections of any rank and any state of the model, and reconstructs from this a unital Jordan homomorphism from the algebra the tests generate into the operators on the cyclic (Gelfand--Naimark--Segal) space of the model's state (Proposition~\ref{prop:bri:eta}); for a general Euclidean Jordan algebra the exclusion holds only for states that give an Albert summand positive weight (Theorem~\ref{thm:bridge}). Albert's theorem says nothing about the length of words, about the Helly number five for families of tests (Corollary~\ref{cor:helly}) or about the sharp-effect realization of Corollary~\ref{cor:ins:sharpalbert}, whose tests are not ideal. For Albert atoms the thirteen traces, and with them $\Acal$, in a normal form a squared octonion associator, are functions of the probabilities of words of length at most four, yet an ideal Hilbert-space model matches every word through length six from $e/3$ (Proposition~\ref{prop:cer:universalsix}), and seven is the first length at which the table excludes every such model (Theorem~\ref{thm:cer:uniform}). The traces exclude a representation only through the purity relations, which involve longer words, and the model of Proposition~\ref{prop:cer:universalsix} reproduces them as averages over two quaternionic blocks whose conditional states differ from the octonionic ones, a difference that first shows as a variance at length seven. A good faith effort was made, using AI systems, to search for prior work; Section~\ref{sec:prior} describes the sources found and how each relates to the results here. Section~\ref{sec:sketch} sketches the proofs, and the external theorems used without proof are listed in Appendix~\ref{app:A:imported}. \section{Sketch of the proofs}\label{sec:sketch} A Hilbert-space model is not asked to embed $\J$, so the proof of Theorem~\ref{thm:bridgeJ} has to show that reproducing all word probabilities forces enough structure inside the model for Albert's theorem to apply. It has three steps, and the four-point case has a second, more explicit route. \paragraph{Matching statistics forces purity relations.} Let $q$ be an atom of $\J$. After the outcome ``yes'' on $q$ the unnormalised state is a multiple of $q$, since the L\"uders update satisfies $U_qx=\ip qx\,q$, with the trace form $\ip qx=\tr(q\jp x)$ (Section~\ref{sec:set:eja}). For any later word $v=(v_1,\dots,v_m)$ put $c=\tr(U_{v_m}\cdots U_{v_1}q)$ and let $\check v=(v_{m-1},\dots,v_1)$ be $v$ without its last letter, in reverse order. The fundamental formula $U_{U_ab}=U_aU_bU_a$ gives, for every history $t$ (a word performed before $q$), \[ \Pr(t,q,v,\check v,q)-2c\,\Pr(t,q,v)+c^2\Pr(t,q)=0 . \] On the model side let $E_v$ be the sequential product of the projections of the outcomes of $v$, $P_t$ the product of the projections of the history, and $\Omega$ the cyclic vector of the cyclic representation of the state, the Hilbert space that the Gelfand--Naimark--Segal construction builds from the words and the state. In a sequential projective representation the same combination of the model's probabilities is the squared norm $\|z\,P_t\Omega\|^2$ with $z=P_qE_vP_q-cP_q$ (Lemmas~\ref{lem:bri:gns} and~\ref{lem:purity}). So matching three word probabilities kills one vector, and matching all words kills the operator $z$ (Lemma~\ref{lem:bri:null}). The identities $z=0$ are the purity relations. \paragraph{The relations rebuild a Jordan homomorphism.} The Peirce reflection $U_{2p_i-e}$, the counterpart of conjugation by $2P_i-1$, is a Jordan automorphism and a signed sum of the identity and the two L\"uders updates of test $i$; on the model side, conjugation by $2P_i-1$ is the same signed sum (equation~\eqref{eq:bri:refl}). So the data control reflected projections as well as the original ones. Images of the atoms under words in the reflections span the algebra $B$ generated by the tests (Lemma~\ref{lem:bri:orbit}). Two families of relations follow: purity relations $P_\alpha P_\beta P_\alpha=\lambda_{\alpha\beta} P_\alpha$ between reflected projections, and every linear relation among reflected atoms in $B$, transported to the projections. In the cyclic representation both families vanish, and the transport $\eta$ from $B$ to operators is a well-defined unital Jordan homomorphism with $\ip\Omega{\eta(x)\Omega}=\ip\rho x$ (Proposition~\ref{prop:bri:eta}). \paragraph{Exceptionality finishes.} If $B=\J$, then $\eta$ is a Jordan homomorphism from $\J$ into the associative algebra of bounded operators. Theorem~\ref{thm:tet:albert} (Appendix~\ref{sec:tet}), an algebraic form of Albert's theorem~\cite{Albert1934}, gives $\eta=0$, against $\eta(e)=1$. Conversely, if the atoms generate a proper subalgebra, it is special and embeds in some $H_D(\C)$. Inside the image the state has a unique trace representative, which is positive, so it is a density matrix that reproduces every word (Lemma~\ref{lem:bri:state}); an order-unit Hahn--Banach extension gives a second proof (Appendix~\ref{app:bri:extend}). In a general Euclidean Jordan algebra the argument needs each test or its complement to be primitive in $B$, since that is what gives the purity relations (Theorem~\ref{thm:bridge}). \paragraph{The tetrahedron.} For four tests the purity relations with words of length at most three read $P_aP_bP_a=o_{ab}P_a$, $P_a(P_bP_c+P_cP_b)P_a=2R_{abc}P_a$ and its three-letter analogue (R3) (Lemma~\ref{lem:fp:access}), with the transition probabilities $o_{ab}$ and the traces $R_{abc}=\ip{p_a\jp p_b}{p_c}$ of Definition~\ref{def:traces}. Fix a vertex $j$ of the tetrahedron of tests and form the three triangle holonomies, such as $h_1=P_jP_kP_lP_j/\sqrt{o_{jk}o_{kl}o_{lj}}$, on the range of $P_j$. The relations make each $h_a$ unitary with scalar Hermitian part $c_a$, the cosine of the triangle's Bargmann phase, and give the skew parts $N_a$ Clifford relations $N_aN_b+N_bN_a=-2M_{ab}$ for a matrix $M$ of traces. The product $h_1h_2h_3$ is, after two contractions, the holonomy of the far face, whose Hermitian part is fixed by a fourth cosine $c_4$. Expanding the product gives $c_4=S+\Gamma$, with $S$ a scalar and $\Gamma$ the volume element of the Clifford relations, and regrouping products, which needs associativity, gives $\Gamma^2=\det M$. Hence $\Dtet=\det M-(c_4-S)^2=0$ in every representation (Theorem~\ref{thm:tetra}). \paragraph{The octonions break it.} In a normal form of four atoms of $\J$, the three triangle phases become unit octonions $h_a=c_a+v_a$. Their imaginary parts satisfy the pairwise anticommutator relations $v_av_b+v_bv_a=-2M_{ab}$, and the real part of $(h_1h_2)h_3$ is the prescribed face scalar $c_4$; only the regrouping fails. By the Harvey--Lawson identity the Gram determinant of three imaginary octonions exceeds their squared triple product by a quarter of their squared associator, and $\Dtet\prod o_{ij}=\Acal$ (Proposition~\ref{prop:Dacal}, Theorem~\ref{thm:fourpoint}). So a positive $\Acal$ excludes every sequential projective representation of the statistics from the maximally mixed state $e/3$ (Corollary~\ref{cor:fp:noideal}). \paragraph{Finitely many words.} Compactness turns the all-word exclusion into a finite linear inequality of unspecified length (Corollary~\ref{cor:bri:finite}). The relations are needed only on the vectors reached from $P_j\Omega$ by at most two holonomies. Recording every word that each use of a relation requires gives the list of $2{,}041$ words of length at most eleven, which decides from the maximally mixed state (Theorem~\ref{thm:eleven}). At length seven the relations hold on enough short histories to fix the Gram matrix of twenty-eight vectors such as $P_AP_zP_yP_x\xi$, where $\xi$ is the normalised state vector after a ``yes'' on $A$, as polynomials in the traces. The same Gram matrix arises in an explicit indefinite form built from the Clifford algebra of the holonomies with one reflection inserted, and there, using once the trace relation (the one polynomial relation among the thirteen traces, Proposition~\ref{prop:fp:tracerel}), three combinations have Gram determinant $-4\Acal^3/\bigl((o_{AB}o_{AC}o_{AD})^9\prod_x(1-o_{Ax})\bigr)<0$. Thus length seven always suffices (Theorem~\ref{thm:cer:uniform}). \section{Ideal tests and their sequences}\label{sec:set} On a Hilbert space the L\"uders rule~\cite{Luders1951} replaces $\rho$ by $P\rho P$ on ``yes'' and by $(1-P)\rho(1-P)$ on ``no'' (unnormalised; the trace is the probability). It leaves unchanged every state on which the outcome is certain. Theorem~\ref{thm:ideal} shows that this property, together with sharpness, forces the same rule in every Euclidean Jordan algebra. \subsection{Euclidean Jordan algebras in brief}\label{sec:set:eja} A \emph{Euclidean} (equivalently, formally real) \emph{Jordan algebra} is a finite-dimensional real vector space $V$ with a commutative bilinear product $x\jp y$, unit $e$, the Jordan identity $(x\jp y)\jp(x\jp x)=x\jp(y\jp(x\jp x))$, and formal reality ($\sum_kx_k\jp x_k=0$ only if all $x_k=0$). An example is $H_n(\C)$ with $x\jp y=\tfrac12(xy+yx)$. A \emph{Jordan homomorphism} between Jordan algebras is a linear map $\phi$ with $\phi(x\jp y)=\phi(x)\jp\phi(y)$, \emph{unital} if it maps unit to unit; a bijective one from $V$ to itself is a \emph{Jordan automorphism}. With the states and effects below it is a Jordan-algebraic system in the sense of~\cite{BarnumWilce2014,BMU2014}. The standard facts below are from Faraut and Kor\'anyi~\cite{FarautKoranyi1994} and are listed with the other imported results in Appendix~\ref{app:A:imported}. \emph{Spectra, trace, cone.} Every $x$ is $\sum_k\lambda_kc_k$ with distinct real $\lambda_k$ and pairwise orthogonal idempotents $c_k$ summing to $e$ \cite[Thm.~III.1.1]{FarautKoranyi1994}. An idempotent is \emph{primitive} if it is nonzero and not a sum of two nonzero orthogonal idempotents; a \emph{Jordan frame} is a set of orthogonal primitive idempotents with sum $e$. Every $x$ is also $\sum_k\mu_kc_k$ over a Jordan frame $\{c_k\}$, the $\mu_k$ not necessarily distinct \cite[Thm.~III.1.2]{FarautKoranyi1994}. The trace $\tr x$ is the sum of the eigenvalues with multiplicity ($\tr q=1$ for primitive $q$), and $\ip xy=\tr(x\jp y)$ is an inner product with $\ip{x\jp y}z=\ip y{x\jp z}$. The cone $V_+$ of squares is closed, self-dual, has $e$ in its interior \cite[Thm.~III.2.1]{FarautKoranyi1994}, and the trace is faithful on it. \emph{Effects} are the $f$ with $0\le f\le e$, \emph{states} the $\rho\ge0$ with $\tr\rho=1$, acting by $f\mapsto\ip\rho f$; every effect counts as a possible outcome. \emph{Quadratic representation and Peirce spaces.} Put $U_ax=2a\jp(a\jp x)-(a\jp a)\jp x$, so $U_ax=axa$ in $H_n(\C)$. Each $U_a$ is self-adjoint and positive, and the fundamental formula $U_{U_ab}=U_aU_bU_a$ holds \cite[Props.~II.3.3, III.2.2]{FarautKoranyi1994}. For an idempotent $p$, $V$ is the orthogonal sum of $V_\lambda(p)=\{x:p\jp x=\lambda x\}$, $\lambda\in\{1,\tfrac12,0\}$, with $V_1\jp V_0=0$, $V_\lambda\jp V_{1/2}\subseteq V_{1/2}$ ($\lambda=0,1$), $V_{1/2}\jp V_{1/2}\subseteq V_1+V_0$ \cite[Prop.~IV.1.1, Thm.~IV.2.1]{FarautKoranyi1994}. For $V=\J$ these spaces are written $\J_\lambda(p)$. Hence $U_{\alpha p+\beta(e-p)}$ acts on $V_1,V_{1/2},V_0$ by $\alpha^2,\alpha\beta,\beta^2$: $U_p$, $U_{e-p}$ are the orthogonal projections onto $V_1(p)$, $V_0(p)$, and the Peirce reflection $U_{2p-e}$ ($+1$ on $V_1\oplus V_0$, $-1$ on $V_{1/2}$) is a Jordan automorphism, the sign being a $\mathbb Z/2$-grading. $V_1(p)$ is a Euclidean Jordan algebra with unit $p$, so a nonzero $p$ is primitive exactly when $V_1(p)=\R p$, and then \begin{equation}\label{eq:set:pure} U_px=\ip px\,p\qquad(p\ \text{primitive},\ x\in V). \end{equation} As a state update, $U_p$ is the \emph{L\"uders map} (a compression in the sense of Alfsen and Shultz~\cite{AlfsenShultz2003}, a filter in the sense of Wilce~\cite{Wilce2019} and of Barnum, M\"uller and Ududec~\cite{BMU2014}). \emph{Classification and speciality.} Unital subalgebras are again Euclidean Jordan algebras; $\Jord\langle e,S\rangle$ is the one generated by $S$. Every Euclidean Jordan algebra is a direct sum of simple ideals, and the simple ones are $H_n(\R)$, $H_n(\C)$, $H_n(\Ha)$, the spin factors $\R\oplus\R^m$ with product $(x_0,\mathbf x)\jp(y_0,\mathbf y)=(x_0y_0+\mathbf x\cdot\mathbf y,\ x_0\mathbf y+y_0\mathbf x)$, and the $27$-dimensional Albert algebra $\J=\HO$ of Hermitian octonionic $3\times3$ matrices with $x\jp y=\frac12(xy+yx)$ (Appendix~\ref{app:A:alb}; Jordan, von Neumann and Wigner~\cite{JordanvNWigner1934}; \cite[Prop.~III.4.4, Thm.~V.3.7]{FarautKoranyi1994}). A Jordan algebra is \emph{special} if it is isomorphic to a Jordan subalgebra of $A^+$, an associative algebra $A$ with product $\tfrac12(xy+yx)$ (for a Euclidean Jordan algebra, special means JC: isomorphic to a Jordan algebra of self-adjoint operators on a complex Hilbert space). All simple ones except $\J$ are special; $\J$ is not (Albert~\cite{Albert1934}). \provisionalref{Atoms}{def:atom}{Same statement}{\leandecl{AlbertChallenge.isAtom_iff_primitive}: trace-one idempotents are the primitive ones.} On $\J$, $\tr e=3$, and $e/3$ is the \emph{maximally mixed state}; the primitive idempotents are the \emph{atoms}, or pure states, and form the octonionic projective plane $\OP$; $o_{ij}=\tr(p_i\jp p_j)\in[0,1]$ is the transition probability between atoms $p_i$ and $p_j$. \provisionalref{Support}{lem:set:support}{Lean scope narrower}{\leandecl{native_positive_zero_pairing_support}. Stated for $V=\J$ only.} \begin{lemma}[Support]\label{lem:set:support} Let $c$ be an idempotent and $x\ge0$ with $\ip xc=0$. Then $U_cx=0$ and $x=U_{e-c}x$. \end{lemma} \begin{proof} $U_cx\ge0$ and $\tr U_cx=\ip cx=0$, so $U_cx=0$. Write $x=x_{1/2}+x_0$. For real $t\ne0$, $U_{tc+(e-c)}x=tx_{1/2}+x_0\ge0$; dividing by $|t|$ and letting $t\to\pm\infty$ gives $\pm x_{1/2}\ge0$, so $x_{1/2}=0$. \end{proof} \subsection{Ideal tests}\label{sec:set:ideal} A two-outcome \emph{instrument} on $V$ is a pair $(T_1,T_0)$ of positive linear maps with $\tr(T_1\rho+T_0\rho)=\tr\rho$; outcome $k$ has probability $\tr T_k\rho$ and leaves $T_k\rho$. Its \emph{measured effect} is $a=T_1^*(e)$, and $F_1=\{\rho:\ip\rho a=1\}$, $F_0=\{\rho:\ip\rho a=0\}$ (states) are the exposed faces on which outcome $1$, resp.\ $0$, is certain. Complete positivity is not assumed. \provisionalref{Ideal test}{def:ideal}{Lean scope narrower}{\leandecl{PositiveInstruments.IsIdealTest}. Defined for finite-dimensional positive forms; the trace-class Hilbert-space version is not defined.} \begin{definition}[Ideal test]\label{def:ideal} A two-outcome instrument $(T_1,T_0)$ is an \emph{ideal test} if (Id) (\emph{ideality}) $T_k\rho=\rho$ for every $\rho\in F_k$, $k=0,1$; and (Sh) (\emph{sharpness}) there are no state $\sigma$ and effect $f$ with $\ip\sigma f=1$ and $\ip\rho f=0$ for all $\rho\in F_0\cup F_1$. The same definition applies to a real, complex or quaternionic Hilbert space $\Hil$ of any dimension, with positive real-linear maps on self-adjoint trace-class operators, density operators as states, effects $0\le f\le1$, and $a=T_1^*(1)$. \end{definition} Condition (Id) is ideality in the sense of Busch, Grabowski and Lahti~\cite{BuschGrabowskiLahti1995}, and G\"unaydin, Piron and Ruegg~\cite{GPR1978} call the octonionic case an ideal measurement of the first kind. As (Id) constrains the instrument only on the faces, (Sh) is needed (Example~\ref{ex:set:c3}). On a Euclidean Jordan algebra, (Sh) holds exactly when the measured effect is an idempotent (proof of Theorem~\ref{thm:ideal}(a)). \provisionalref{Ideal tests}{thm:ideal}{Lean scope narrower}{\leandecl{PositiveInstruments.idealTest_iff_quadraticRepresentation}. Part (a) only.} \begin{theorem}[Ideal tests are L\"uders instruments]\label{thm:ideal} \leavevmode \begin{enumerate}[label=(\alph*),leftmargin=2em] \item Let $V$ be a Euclidean Jordan algebra. A two-outcome instrument on $V$ is an ideal test if and only if $T_1=U_p$ and $T_0=U_{e-p}$ for an idempotent $p\in V$; then $p$ is the measured effect. \item Let $\Hil$ be a real, complex or quaternionic Hilbert space of any dimension. A two-outcome instrument on $\Hil$ is an ideal test if and only if $T_1X=PXP$ and $T_0X=(1-P)X(1-P)$ for an orthogonal projection $P$; then $P$ is the measured effect. \end{enumerate} \end{theorem} Theorem~\ref{thm:ideal} is a variant, with these operational hypotheses, of known characterisations: on a unital JB-algebra the pairs of bicomplementary positive projections are exactly the pairs $(U_p,U_{e-p})$ (Alfsen and Shultz~\cite[Thms.~2.83, 7.12]{AlfsenShultz2003}, with hypotheses on kernels and ranges); on $\BH$, ideality implies the L\"uders rule under the hypotheses of Kleinmann~\cite[Sec.~III\,A]{Kleinmann2014}, Chiribella and Yuan~\cite{ChiribellaYuan2018} and Cho~\cite[Thm.~6.3.34]{Cho2019}. These sources treat operator algebras; reconstructions of Jordan-algebraic systems take $U_p$ as an axiom~\cite{BMU2014}. \begin{proof}[Proof of (a)] \emph{The measured effect is an idempotent.} For $\rho\ge0$, $\ip\rho a=\tr T_1\rho\ge0$ and $\ip\rho{e-a}=\tr T_0\rho\ge0$, so $0\le a\le e$ by self-duality. Let $a=\sum_\lambda\lambda c_\lambda$ and $p=c_1$ ($p$ and $c_0$ possibly $0$). For a state $\rho$, $\ip\rho{e-a}=\sum_{\lambda<1}(1-\lambda)\ip\rho{c_\lambda}$ is a sum of nonnegative terms, so $\rho\in F_1$ iff $\ip\rho{e-p}=0$, iff $\rho\in V_1(p)$ by Lemma~\ref{lem:set:support}; likewise $F_0$ is the set of states in $V_1(c_0)$. If $r=e-p-c_0\neq0$, the state $r/\tr r$ and the effect $r$ violate (Sh), since $r\in V_0(p)\cap V_0(c_0)$ is orthogonal to every state in $V_1(p)$ or $V_1(c_0)$. So $r=0$ and $a=p$. \emph{$T_1$ on the Peirce spaces of $p$.} By (Id), $T_1$ fixes the states in $V_1(p)$, which span $V_1(p)$. For $y\in V_0(p)$, $y\ge0$: $T_1y\ge0$ and $\tr T_1y=\ip yp=0$, so $T_1y=0$; such $y$ span $V_0(p)$. For $z\in V_{1/2}(p)$ choose $\kappa>0$ with $e+z/\kappa\ge0$. For every real $t$, \[ 0\le U_{p+t(e-p)}\bigl(e+z/\kappa\bigr)=p+\tfrac t\kappa\,z+t^2(e-p), \] and applying $T_1$ gives $p+\tfrac t\kappa T_1z\ge0$; letting $t\to\pm\infty$, $T_1z=0$. So $T_1=U_p$. The instrument $(T_0,T_1)$ satisfies the same hypotheses with measured effect $e-p$, so $T_0=U_{e-p}$. \emph{Converse.} $U_p,U_{e-p}$ are positive, $U_pe+U_{e-p}e=e$, and they fix the states of $V_1(p)$ and $V_0(p)$, which are $F_1$ and $F_0$; so (Id) holds. An effect vanishing on these states has $\ip pf=\ip{e-p}f=0$, so $\tr f=0$ and $f=0$; so (Sh) holds. \end{proof} Part~(b) is the same argument with spectral projections (Appendix~\ref{app:bri:ideal}). It gives the L\"uders, not the von Neumann, rule for degenerate $P$, and repeatability, that an immediate repetition of the test confirms its outcome, follows from $U_p^2=U_p$. \subsection{Words and their probabilities}\label{sec:set:seq} Fix $V$, idempotents $p_1,\dots,p_n$ (the \emph{tests}; a test is identified with its idempotent) and a state $\rho$. Test $i$ has outcome \emph{yes}$_i$ with update $U_{p_i}$ and \emph{no}$_i$ with update $U_{e-p_i}$. A \emph{word} (an outcome sequence) is a finite sequence $w=(q_1,\dots,q_m)$, $q_t\in\{p_i,e-p_i\}$, in chronological order, with probability \begin{equation}\label{eq:set:prob} \Pr\nolimits_\rho(w)=\tr\bigl(U_{q_m}\cdots U_{q_1}\rho\bigr)=\ip\rho{\theta(w)}, \qquad \theta(w)=U_{q_1}U_{q_2}\cdots U_{q_m}e . \end{equation} The effect $\theta(w)$ is the iterated sequential product of the outcomes in the sense of Gudder and Greechie~\cite{GudderGreechie2002} (on a Euclidean Jordan algebra an invariant continuous sequential product is $U_{\sqrt a}b$~\cite{vandeWetering}); the empty word has $\theta=e$. A skipped test is absent from the word. Summing over the outcomes of the last test gives the shorter word, but summing over a test in the middle does not, since $U_p+U_{e-p}$ removes the $V_{1/2}(p)$ component. A \emph{table} is a finite labelled family of tests together with all its word probabilities. Every $\theta(w)$ lies in $B=\Jord\langle e,p_1,\dots,p_n\rangle$, so these sequential statistics (also called temporal correlations~\cite{BMKG2013,VitaglianoBudroni2022}) depend only on $B$, the tests, and $\rho|_B$. If all letters are atoms, \eqref{eq:set:pure} gives $\Pr_\rho(w)=\ip\rho{q_1}\prod_t\ip{q_t}{q_{t+1}}$. On a Hilbert space, projections $P_i$ and the L\"uders rule give \[ \Pr(w)=\Tr(Q_m\cdots Q_1\varrho Q_1\cdots Q_m)=\Tr(\varrho E_w), \] with \[ E_w=Q_1\cdots Q_{m-1}Q_mQ_{m-1}\cdots Q_1,\qquad Q_t=P_i\ (\text{yes}_i),\quad Q_t=1-P_i\ (\text{no}_i), \] the counterpart of $\theta(w)$, since $U_QX=QXQ$. \paragraph{Other measurements.}\label{sec:set:dict} The test of $e-p$ is that of $p$ with outcomes relabelled. A \emph{complete measurement} is the three-outcome L\"uders instrument $(U_{f_1},U_{f_2},U_{f_3})$ of a Jordan frame of $\J$; reading only whether the outcome was $k$ updates by $\sum_{l\ne k}U_{f_l}\ne U_{e-f_k}$, so it is not a family of tests (Proposition~\ref{prop:markov}). A \emph{generalised L\"uders instrument} has non-idempotent effects $a_k$ and the square-root update $x\mapsto U_{a_k^{1/2}}x$~\cite{BuschSingh1998}. \subsection{Three model classes}\label{sec:set:comp} Every theorem names one of the following three model classes; class (I) is the comparison class of every main theorem. \provisionalref{Class (I)}{def:idealmodel}{Same statement}{\leandecl{HilbertProjectiveData} ($\R$, $\C$), \leandecl{Quaternionic}\allowbreak\leandecl{ProjectiveData}; the proofs use the wider class \leandecl{AlgebraicProjectiveData}.} \begin{definition}[Sequential projective representation; class (I)]\label{def:idealmodel} Let $V$, tests $p_1,\dots,p_n$ and a state $\rho$ be given. A \emph{sequential projective representation} of their statistics consists of a Hilbert space $\Hil$ over $\R$, $\C$ or $\Ha$ of arbitrary (possibly infinite, possibly nonseparable) dimension, one orthogonal projection $P_i$ on $\Hil$ for each test, of arbitrary rank, and a state $\varphi$ of the representation, that is, a positive real-linear functional on the bounded self-adjoint operators with $\varphi(1)=1$ (not necessarily normal, faithful or pure), such that \begin{equation}\label{eq:set:match} \varphi(E_w)=\Pr\nolimits_\rho(w)\qquad\text{for every word }w . \end{equation} If \eqref{eq:set:match} is required only for a given set of words, it is a sequential projective representation \emph{of those words}. \end{definition} The same projection is used each time a test is performed, and both outcomes update by the L\"uders rule. For complex spaces and density-operator states these are the sequential projective quantum representations of Budroni, Moroder, Kleinmann and G\"uhne~\cite{BMKG2013}, who showed that membership on a finite set of words is decided, in arbitrary dimension, by a moment-matrix condition. Real and quaternionic spaces and non-normal states give the same data (Lemma~\ref{lem:bri:gns} and Appendix~\ref{app:bri:field}), so their name is used here. A degeneracy-breaking rule is not in the class; with such rules dimension-dependent bounds appear~\cite{BudroniEmary2014}. \begin{remark}[A fixed unitary between tests]\label{rem:set:hidden} Let $\Hil$ be finite-dimensional, with projections $P_i$, a density operator $\varrho$ and a unitary $Y$ applied before each test, so that a word has probability $\Tr(Q_mY\cdots Q_1Y\varrho Y^*Q_1\cdots Y^*Q_m)$. If these probabilities equal $\Pr_\rho(w)$ for every word, the table has a sequential projective representation without $Y$. For a word $u$ let $\sigma_u$ be the unnormalised operator that meets the next test after $u$ ($\sigma_\varnothing =Y\varrho Y^*$), and $K$ the span of the ranges of all $\sigma_u$. In every Euclidean Jordan algebra $U_{e-p}U_p=0$, so a ``yes'' and then a ``no'' on the same test, or the reverse, has probability zero after every $u$. In the model the first is $\|(1-P_i)YP_i\sigma_u^{1/2}\|_2^2$ (Hilbert--Schmidt norm), so $(1-P_i)YP_i$ and likewise $P_iY(1-P_i)$ vanish on $K$, that is $P_iY=YP_i$ on $K$. The range of $\sigma_{(u,q)}$ is $YQ$ applied to that of $\sigma_u$, so $YP_iK\subseteq K$ and $Y(1-P_i)K\subseteq K$; hence $YK\subseteq K$, so $YK=K$ in finite dimension, and $P_iK=Y^{-1}YP_iK\subseteq K$. On $K$ the unitary commutes with every $Q_t$, so $Q_mY\cdots YQ_1=Y^{m-1}Q_m\cdots Q_1$ there, and invariance of the trace gives the table of $(K,P_i|_K,\sigma_\varnothing)$, a sequential projective representation. The step $YK=K$ uses finite dimension; infinite dimension is not treated here. \end{remark} \provisionalref{Class (II)}{def:memorymodel}{Same statement}{\leandecl{IsNormalizedFinite}\allowbreak\leandecl{CPModel}, \leandecl{RealFiniteKrausModel}.} \begin{definition}[Completely positive realization; class (II)]\label{def:memorymodel} Let a finite family of instruments be given, with one letter per outcome. A \emph{completely positive realization} of dimension $d$, in the sense of Monr\`as and Winter~\cite{MonrasWinter2016} (also called a finite quantum memory), is a density matrix $\sigma$ on $\C^d$ and one completely positive map $\mathcal E_a$ on $M_d(\C)$ per letter, the same each time the letter occurs, such that the outcome maps of each instrument sum to a trace-preserving map and $\Tr(\mathcal E_{a_m}\cdots\mathcal E_{a_1}\sigma)=\Pr(a_1\cdots a_m)$ for every word. For tests the letters are $p_i,e-p_i$, the right side is $\Pr_\rho(w)$, and a skipped test is absent from the word. \end{definition} A finite-dimensional complex sequential projective representation is a completely positive realization ($X\mapsto P_iXP_i$, $X\mapsto(1-P_i)X(1-P_i)$), and a real or quaternionic one becomes one after passing to $\C$ (Proposition~\ref{prop:field}(a)). An infinite-dimensional one is not, and a completely positive realization need not use projections or L\"uders updates. So excluding one class does not exclude the other: the four-atom statistics excluded from class (I) have a completely positive realization on $\C^{69}$ (Theorem~\ref{thm:cp69}). \emph{Class (III): arbitrary instruments in infinite dimension.} A model always exists: a classical register stores the current normalised state of $V$ and emits each outcome with its conditional probability. Statements about class (III) are constructions, never exclusions. \paragraph{Terminology.} Terms defined later: \emph{$B$-primitive} tests (Definition~\ref{def:bri:minimal}); the \emph{purity operator} and \emph{purity relations} (Lemma~\ref{lem:purity}); \emph{orbit labels} (Section~\ref{sec:bri:recon}); the \emph{four-point invariant} $\Acal$ (\eqref{eq:A:Acal52}) and the \emph{double-pair traces} (Definition~\ref{def:traces}); the \emph{tetrahedron identity} and the \emph{tetrahedron statistic} $\Dtet$ (Theorem~\ref{thm:tetra}); the \emph{triangle holonomies} (Section~\ref{sec:fp:tetra}); the \emph{minimum separating word length} (Section~\ref{sec:cer}); \emph{projective equalities} (Section~\ref{sec:cer:equality}); and \emph{line subalgebras} (Section~\ref{sec:fiv:helly}). \section{A sequential criterion for speciality}\label{sec:bri} When do the word probabilities of Section~\ref{sec:set} have a sequential projective representation (Definition~\ref{def:idealmodel})? For atoms of the Albert algebra the answer is: exactly when the atoms generate a proper subalgebra of $\J$. \subsection{Statement for the Albert algebra}\label{sec:bri:statements} \provisionalref{Albert criterion}{thm:bridgeJ}{Different form}{\leandecl{AlbertChallenge.albert_criterion}: (1) as an embedding of a subalgebra containing the atoms into complex Hermitian matrices; (2) with $d\in\{6,24\}$ and projections fixed for all states; (3) on spaces in Lean's lowest universe, algebraic models admitted. (3) in any universe, and the last two sentences: \leandecl{AlbertChallenge.no_representation_of_generates}, \leandecl{AlbertChallenge.linear_separator_of_generates}.} \begin{theorem}[The Albert algebra]\label{thm:bridgeJ} Let $p_1,\dots,p_n$ be atoms of $\J=\HO$ and $B=\Jord\langle e,p_1,\dots,p_n\rangle$. The model class is (I). The following are equivalent. \begin{enumerate}[label=(\arabic*),leftmargin=2em] \item $B$ is special; equivalently, $B\neq\J$. \item For every state $\rho$ of $\J$ there is a sequential projective representation on $\C^d$ for some finite $d$, with a density-matrix state. \item For at least one state $\rho$ of $\J$ there is a sequential projective representation, over $\R$, $\C$ or $\Ha$, of any dimension, with projections of any rank and any state $\varphi$. \end{enumerate} If $B=\J$ there is no sequential projective representation from any state. Moreover there are then a length $L$ and a linear inequality in the probabilities of words of length at most $L$ that holds in every sequential projective representation and is violated by the statistics of $p_1,\dots,p_n$ from every state of $\J$. \end{theorem} The same holds if some tests are complements $e-p$ of atoms, since the test of $e-p$ is the test of $p$ with outcomes relabelled and $B$ does not change. Four atoms can generate $\J$ (Theorem~\ref{thm:fourpoint} gives an exact criterion), while three never do (Proposition~\ref{prop:three}). \begin{remark}[Nothing non-associative happens in time]\label{rem:bri:assoc} The update maps $U_{p_i}$ and $U_{e-p_i}$ are linear maps of the $27$-dimensional real vector space $\J$. They compose associatively, and the probability \eqref{eq:set:prob} of a word is obtained by composing them in the order of the outcomes, so no ``non-associativity in time'' is observed. The non-associative Jordan product enters only through \emph{which} linear maps the updates are (for a model without the L\"uders requirement, see Theorem~\ref{thm:cp69}). \end{remark} \subsection{The mechanism: null vectors on every history}\label{sec:bri:null} Fix a sequential projective representation $(\Hil,P_i,\varphi)$. The first step keeps only what the statistics can see. \provisionalref{Cyclic representation}{lem:bri:gns}{Same statement}{\leandecl{cyclicHilbertModel}, \leandecl{cyclicHilbertModel_probability}, starting from an algebraic model.} \begin{lemma}[Real cyclic representation]\label{lem:bri:gns} Let $A_0$ be the real $*$-algebra spanned by the finite products of $P_1,\dots,P_n$ (the empty product $1$ included). Put $\varphi_s(X)=\varphi\bigl(\tfrac12(X+X^*)\bigr)$ and $\ip XY_\varphi=\varphi_s(X^*Y)$. This form is symmetric and positive semidefinite; let $\Hil_\varphi$ be the real Hilbert space obtained by quotienting its null space and completing, and $\Omega=[1]$. Left multiplication by $X\in A_0$ descends to a bounded operator $\pi(X)$ on $\Hil_\varphi$, $\pi$ is a real $*$-representation, and each $\pi(P_i)$ is an orthogonal projection. For a word $w=(q_1,\dots,q_m)$ put $P_w=Q_m\cdots Q_1$ (last outcome leftmost). Then \[ \Pr(w)=\varphi(E_w)=\bigl\|\pi(P_w)\Omega\bigr\|^2, \] and the vectors $\pi(P_w)\Omega$, $w$ ranging over all words, span a dense subspace of $\Hil_\varphi$. \end{lemma} \begin{proof} This is the Gelfand--Naimark--Segal construction, written so that $\varphi$ is evaluated only on self-adjoint operators; it therefore covers real and quaternionic spaces and non-normal or non-faithful states. Since each $Q_t$ is a self-adjoint idempotent, $E_w=P_w^*P_w$, so $\varphi(E_w)=\varphi_s(P_w^*P_w)=\|\pi(P_w)\Omega\|^2$. Every finite product of the $P_i$ is $P_w$ for an all-``yes'' word $w$, which gives density. Boundedness and the $*$-property are routine (Appendix~\ref{app:bri:gns}). \end{proof} The data enter through squares. A \emph{formal word combination} is a finite real combination $X=\sum_wc_w[w]$ of word symbols. It has two evaluations: the operator $E(X)=\sum_wc_wE_w\in A_0$ in the representation, and the effect $\theta(X)=\sum_wc_w\theta(w)\in B$ in $V$. Matching all words means exactly $\varphi(E(X))=\ip\rho{\theta(X)}$ for every $X$. Prefixing a word $t=(t_1,\dots,t_k)$ to every word of $X$ gives $t\cdot X$, and \begin{equation}\label{eq:bri:prefix} E(t\cdot X)=P_t^*\,E(X)\,P_t,\qquad \theta(t\cdot X)=U_{t_1}\cdots U_{t_k}\,\theta(X). \end{equation} \provisionalref{Null vectors}{lem:bri:null}{Lean scope narrower}{\leandecl{allHistoryNullContract}: the null-set part only, left invariance of the null vectors and extension from word vectors to their span; the step from $E(X)=Z^2$, $\theta(X)=0$ and matching to $\pi(Z)\pi(P_t)\Omega=0$ is not part of this declaration.} \begin{lemma}[Vanishing moments give null vectors on every history]\label{lem:bri:null} Let $Z\in A_0$ be self-adjoint and $X$ a formal word combination with $E(X)=Z^2$ as an identity of operators and $\theta(X)=0$. Let $\mathcal T$ be a set of words (histories). If the representation reproduces $\Pr(t\,w)$ for every $t\in\mathcal T$ and every $w$ in the support of $X$, then $\pi(Z)\pi(P_t)\Omega=0$ for every $t\in\mathcal T$. If it reproduces all words, then $\pi(Z)=0$. \end{lemma} \begin{proof} By Lemma~\ref{lem:bri:gns}, \eqref{eq:bri:prefix} and matching, \[ \begin{aligned} \|\pi(Z)\pi(P_t)\Omega\|^2&=\varphi(P_t^*Z^2P_t)=\varphi(E(t\cdot X))\\ &=\sum_wc_w\Pr(t\,w)=\ip\rho{U_{t_1}\cdots U_{t_k}\theta(X)}=0 . \end{aligned} \] If all words are reproduced, the bounded operator $\pi(Z)$ vanishes on a dense subspace. \end{proof} The hypothesis puts $ZP_t$ in the null left ideal of the moment functional, in the language of Pironio, Navascu\'es and Ac\'{\i}n~\cite{PNA2010} (Section~\ref{sec:prior:temporal}). The quantifier over histories matters: when $\varphi$ is not faithful, one vanishing expectation says nothing about the operator, and finite arguments (Section~\ref{sec:cer}) must record which histories they use. Vanishing effects $\theta(X)=0$ come from purity: a ``yes'' on an atom leaves a multiple of that atom, and in the algebra this is the identity \begin{equation}\label{eq:bri:lambda} U_qx=\lambda_q(x)\,q,\qquad \lambda_q(x)=\ip qx/\ip qq\qquad(x\in B), \end{equation} valid for every idempotent $q$ that is primitive in $B$, because then $U_qB=B_1(q)=\R q$. For an atom $q$ of $\J$, $\ip qq=1$ and \eqref{eq:bri:lambda} is \eqref{eq:set:pure}. \provisionalref{Purity}{lem:purity}{Lean scope narrower}{\leandecl{purityResidual_threeHistories}; the vanishing in $V$, for outcomes that are atoms of $\J$: \leandecl{NativeAnalysis.native_corner_long_probability}.} \begin{lemma}[Purity]\label{lem:purity} Let $q$ be one of the outcomes $p_i$ or $e-p_i$, primitive in $B$ (for instance an atom of $\J$), and $P_q$ the corresponding projection $P_i$ or $1-P_i$. Let $v=(v_1,\dots,v_m)$ be a word, $c=\lambda_q(\theta(v))$ (for an atom, $c=\tr(U_{v_m}\cdots U_{v_1}q)$), and \[ z_q(v)=P_q\,E_v\,P_q-c\,P_q . \] For every word $t$, \begin{equation}\label{eq:bri:purity} \bigl\|\pi(z_q(v))\pi(P_t)\Omega\bigr\|^2=\Pr(t,q,v,\check v,q)-2c\,\Pr(t,q,v)+c^2\Pr(t,q), \end{equation} with the probabilities of the representation on the right and $\check v=(v_{m-1},\dots,v_1)$; the same combination of probabilities in $V$ is zero. Hence if the representation reproduces these three words, $\pi(z_q(v))\pi(P_t)\Omega=0$; if it reproduces all words, $\pi(z_q(v))=0$. \end{lemma} The operator $z_q(v)$ is called the purity operator and $\pi(z_q(v))=0$ the purity relations. \begin{proof} Expanding, $z_q(v)^2=P_qE_vP_qE_vP_q-2cP_qE_vP_q+c^2P_q$, and these three operators are the sequential products $E$ of the words $(q,v,\check v,q)$, $(q,v)$ and $(q)$; so $z_q(v)^2=E(X)$ for $X=[q,v,\check v,q]-2c[q,v]+c^2[q]$, and \eqref{eq:bri:purity} is Lemma~\ref{lem:bri:null}'s computation. \emph{The effect vanishes.} In $V$, iterating the fundamental formula gives \[ U_{v_1}\cdots U_{v_{m-1}}U_{v_m}U_{v_{m-1}}\cdots U_{v_1}=U_{\theta(v)}, \] hence, by \eqref{eq:bri:lambda}, \[ \theta(q,v,\check v,q)=U_qU_{\theta(v)}U_qe=U_{U_q\theta(v)}e=U_{cq}e=c^2q, \qquad\theta(q,v)=U_q\theta(v)=c\,q, \] so $\theta(X)=c^2q-2c^2q+c^2q=0$. For an atom, $c=\ip q{\theta(v)}=\tr(U_{v_m}\cdots U_{v_1}q)$ because the $U$'s are self-adjoint. No conditional probability is divided by a possibly zero number. \end{proof} Lemma~\ref{lem:purity} feeds the certificates of Section~\ref{sec:cer}; the reconstruction needs it only for $v$ of length one, transported along reflections. \subsection{Reconstruction of a Jordan homomorphism}\label{sec:bri:recon} Assume now that a sequential projective representation reproduces all words, and that every test $p_i$ is primitive in $B$; for atoms of $\J$ with $B=\J$ this holds. The construction builds a unital Jordan homomorphism from $B$ into the bounded self-adjoint operators $B(\Hil_\varphi)_{\sa}$ out of operators the data control, using no property of the tests beyond \eqref{eq:bri:lambda}. \emph{Reflections are controlled by the data.} Put $r_i=2p_i-e$ and $\Sigma_i=2P_i-1$. The Peirce reflection $U_{r_i}$ is a Jordan automorphism of $B$ (Section~\ref{sec:set:eja}), and it is a combination of the two L\"uders updates; so is conjugation by $\Sigma_i$: \begin{equation}\label{eq:bri:refl} U_{r_i}=2(U_{p_i}+U_{e-p_i})-\id,\qquad \Sigma_iY\Sigma_i=2P_iYP_i+2(1-P_i)Y(1-P_i)-Y . \end{equation} \emph{Orbit labels.} An \emph{orbit label} is a pair $\alpha=(g,i)$ with $g=(g_1,\dots,g_k)$ a finite sequence of indices; it names \[ q_\alpha=U_{r_{g_1}}\cdots U_{r_{g_k}}\,p_i\in B,\qquad P_\alpha=\Sigma_{g_1}\cdots \Sigma_{g_k}\,P_i\,\Sigma_{g_k}\cdots \Sigma_{g_1}\in A_0 . \] Each \emph{orbit point} $q_\alpha$ is primitive in $B$, as the image of a primitive idempotent under an automorphism, and $P_\alpha$ is a projection. Write $r_\alpha=2q_\alpha-e$ and $\Sigma_\alpha=2P_\alpha-1$. Different labels may name the same $q_\alpha$ but, a priori, different $P_\alpha$, until the linear relations below identify them. Composing the conjugations by $\Sigma_j$ gives (Appendix~\ref{app:bri:lifts}): \provisionalref{Lifts}{lem:bri:lift}{Same statement}{\leandecl{hasSequentialEffectLift_orbit_corner}, \leandecl{hasSequentialEffectLift_reflection}.} \begin{lemma}[Lifts of conjugated products]\label{lem:bri:lift} For orbit labels $\alpha_1,\dots,\alpha_s$ there is a formal word combination $X$, a \emph{lift}, with \[ E(X)=P_{\alpha_1}\cdots P_{\alpha_{s-1}}P_{\alpha_s}P_{\alpha_{s-1}}\cdots P_{\alpha_1}, \qquad\theta(X)=U_{q_{\alpha_1}}\cdots U_{q_{\alpha_{s-1}}}q_{\alpha_s}. \] Moreover, if $E(X)=Y$ and $\theta(X)=y$, there is $X'$ with $E(X')=\Sigma_\alpha Y\Sigma_\alpha$ and $\theta(X')=U_{r_\alpha}y$. These identities hold for any projections $P_i$. \end{lemma} \emph{The orbit spans $B$.} The Peirce decomposition gives, for an idempotent $q$ and any $x$, and in any associative algebra for a projection $P$ and any $Y$, \begin{equation}\label{eq:bri:peirceid} q\jp x=U_qx+\tfrac14\bigl(x-U_{r_q}x\bigr)\quad(r_q=2q-e), \end{equation} \begin{equation}\label{eq:bri:assocpeirce} P\jp Y=PYP+\tfrac14\bigl(Y-\Sigma Y\Sigma\bigr)\quad(\Sigma=2P-1), \end{equation} where $P\jp Y=\tfrac12(PY+YP)$. (Both sides of \eqref{eq:bri:peirceid} agree on each Peirce space of $q$.) \provisionalref{Orbit span}{lem:bri:orbit}{Lean scope narrower}{\leandecl{nativeOrbitSpan_iff_generated}, for tests that are atoms of $\J$.} \begin{lemma}[Orbit span]\label{lem:bri:orbit} $B=\spn\bigl(\{e\}\cup\{q_\alpha\}\bigr)$. \end{lemma} \begin{proof} The span $B'$ of the right side lies in $B$ and is mapped into itself by each $U_{q_\alpha}$ (by \eqref{eq:bri:lambda}) and each $U_{r_\alpha}$ (which fixes $e$ and permutes orbit points), hence by each $q_\alpha\jp{}$ (by \eqref{eq:bri:peirceid}). So $B'$ is a unital subalgebra containing every $p_i$, and $B'=B$. Details are in Appendix~\ref{app:bri:orbit}. \end{proof} \emph{Two families of null operators.} Let $\lambda_{\alpha\beta}= \lambda_{q_\alpha}(q_\beta)$, so $U_{q_\alpha}q_\beta=\lambda_{\alpha\beta}q_\alpha$. \begin{itemize}[leftmargin=1.5em] \item \emph{Purity relations} (transported). $T_{\alpha\beta}=P_\alpha P_\beta P_\alpha-\lambda_{\alpha\beta}P_\alpha$. By Lemma~\ref{lem:bri:lift}, $T_{\alpha\beta}^2$ has a lift with effect $U_{q_\alpha}U_{q_\beta}q_\alpha-2\lambda_{\alpha\beta}U_{q_\alpha}q_\beta+ \lambda_{\alpha\beta}^2q_\alpha$ (Appendix~\ref{app:bri:lifts}), and this is $0$ because the fundamental formula gives $U_{q_\alpha}U_{q_\beta}q_\alpha=U_{q_\alpha}U_{q_\beta}U_{q_\alpha}e= U_{\lambda_{\alpha\beta}q_\alpha}e=\lambda_{\alpha\beta}^2q_\alpha$. For $\alpha=(\varnothing,i)$, $\beta=(\varnothing,j)$ this is Lemma~\ref{lem:purity} with $q=p_i$, $v=(p_j)$. \item \emph{Linear relations} (transported). Whenever $c_0e+\sum_rc_rq_{\alpha_r}=0$ in $B$, put $Z=c_0\,1+\sum_rc_rP_{\alpha_r}$. Writing each $P_{\alpha_r}\jp P_{\alpha_s}$ in $Z^2$ by \eqref{eq:bri:assocpeirce} and applying Lemma~\ref{lem:bri:lift} gives a lift of $Z^2$ whose effect is, by \eqref{eq:bri:peirceid}, $\bigl(c_0e+\sum_rc_rq_{\alpha_r}\bigr)^2=0$ (Appendix~\ref{app:bri:lifts}). \end{itemize} By Lemma~\ref{lem:bri:null}, $\pi(T_{\alpha\beta})=0$ and $\pi(Z)=0$ for all such operators. In particular (a linear relation with $c_0=0$) $\pi(P_\alpha)=\pi(P_\beta)$ whenever $q_\alpha=q_\beta$. \provisionalref{Reconstruction}{prop:bri:eta}{Lean scope narrower}{\leandecl{exists_generatedCyclicJordanMap}. Stated for tests that are atoms of $\J$, with $\eta$ valued in operators on the cyclic space of an algebraic model.} \begin{proposition}[The reconstructed homomorphism]\label{prop:bri:eta} Suppose a sequential projective representation reproduces all words and every test is primitive in $B$. Then \[ \eta\Bigl(c_0e+\sum_rc_rq_{\alpha_r}\Bigr)=c_0\,1+\sum_rc_r\,\pi(P_{\alpha_r}) \] defines a unital Jordan homomorphism $\eta:B\to B(\Hil_\varphi)_{\sa}$ with $\ip\Omega{\eta(x)\Omega}=\ip\rho x$ for all $x\in B$. \end{proposition} \begin{proof} \emph{Well defined.} By Lemma~\ref{lem:bri:orbit} every $x\in B$ has such an expression, and two expressions differ by a linear relation, which the transported linear relations send to $0$: so $\eta$ is well defined, linear and unital. \emph{Reflections and corners.} It intertwines reflections, $\eta(U_{r_\alpha}x)=\pi(\Sigma_\alpha)\eta(x)\pi(\Sigma_\alpha)$, because $U_{r_\alpha}$ acts on labels as conjugation by $\Sigma_\alpha$ acts on the $P$'s, and the purity relations give the corner relation $\pi(P_\alpha)\eta(x)\pi(P_\alpha)=\lambda_{q_\alpha}(x)\pi(P_\alpha)$ (both checked in Appendix~\ref{app:bri:eta}). \emph{Products.} Now \eqref{eq:bri:peirceid}, \eqref{eq:bri:lambda} and \eqref{eq:bri:assocpeirce} give the decisive line: \[ \eta(q_\alpha\jp x)=\lambda_{q_\alpha}(x)\,\pi(P_\alpha)+\tfrac14\bigl(\eta(x)- \pi(\Sigma_\alpha)\eta(x)\pi(\Sigma_\alpha)\bigr)=\pi(P_\alpha)\jp\eta(x). \] Both sides of $\eta(y\jp x)=\eta(y)\jp\eta(x)$ are linear in $y$, and $B$ is spanned by $e$ and the $q_\alpha$; so $\eta$ is a Jordan homomorphism. Finally $\ip\Omega{\pi(P_\alpha)\Omega}=\varphi(E(X))=\ip\rho{q_\alpha}$ for the lift $X$ of $P_\alpha$ (Lemma~\ref{lem:bri:lift}), by matching. \end{proof} \emph{The contradiction.} Let $B=\J$. The map $\eta$ is a Jordan homomorphism into the real associative algebra of bounded operators on $\Hil_\varphi$. Theorem~\ref{thm:tet:albert} gives $\eta=0$, whereas $\eta(e)=1\ne0$ because $\|\Omega\|=1$; equivalently, since $\J$ is simple, $\eta$ would be a faithful representation of $\J$ and its Jordan product by operators. This contradiction uses the algebraic proof in Appendix~\ref{sec:tet}; Albert's original theorem remains the historical source of exceptionality~\cite{Albert1934}. The argument used no four-point invariant, no finite rank, no normality or faithfulness of $\varphi$, no positivity of $\rho$, and nothing about the original Hilbert space beyond its cyclic representation. \subsection{The converse: discard, embed, construct}\label{sec:bri:converse} Let $\rho$ be a state of $V$ and suppose $B=\Jord\langle e,p_1,\dots,p_n\rangle$ has no summand isomorphic to $\J$ given positive weight by $\rho$ (Section~\ref{sec:bri:general} makes this precise). For atoms of $\J$ with $B\neq\J$ there is nothing to discard: $B$ has dimension below $27$, so by the classification all its simple summands are special. \emph{Discard.} If $A\cong\J$ is a summand of $B$ with unit $c$ and $\ip\rho c=0$, then $|\ip\rho x|\le\|x\|\,\ip\rho c=0$ for $x\in A$ (Appendix~\ref{app:bri:extend}). So $\Pr_\rho(w)=\ip\rho{\theta_{\mathrm{sp}}(w)}$, where $\theta_{\mathrm{sp}}(w)$ is the image of $\theta(w)$ under the central projection $\pi_{\mathrm{sp}}$ of $B$ onto the sum $B_{\mathrm{sp}}$ of its special summands; $\pi_{\mathrm{sp}}$ is a unital Jordan homomorphism, so $\theta_{\mathrm{sp}}(w)$ is the sequential product of the tests $\pi_{\mathrm{sp}}(p_i)$. \emph{Embed.} Each special simple summand has a faithful unital representation on some $\C^{d_j}$, and their direct sum is a faithful unital Jordan homomorphism $\iota:B_{\mathrm{sp}}\to H_d(\C)$ with $x\ge0$ iff $\iota(x)\ge0$ (Appendix~\ref{app:bri:extend}). \emph{Construct the state.} The following lemma gives the density matrix directly. \provisionalref{Explicit state}{lem:bri:state}{Lean scope narrower}{\leandecl{FaithfulHermitian}\allowbreak\leandecl{JordanEmbedding.exists_unique_positive_trace_representative}, \leandecl{suppliedEmbedding}\allowbreak\leandecl{StateTransport}. The density matrix is obtained, as printed, from a positive functional with value $1$ at the unit; the word transport is stated only for unital subalgebras of $\J$ and states of $\J$.} \begin{lemma}[Explicit state]\label{lem:bri:state} Let $W$ be a Euclidean Jordan algebra with unit $e_W$, $\omega$ a linear functional on $W$ with $\omega(x^2)\ge0$ for all $x\in W$ and $\omega(e_W)=1$, $p_1,\dots,p_n$ idempotents of $W$, and $\iota:W\to H_d(\C)$ an injective unital Jordan homomorphism. There is a unique $z\in W$ with \[ \Tr\bigl(\iota(z)\iota(x)\bigr)=\omega(x)\qquad\text{for all }x\in W . \] Then $\sigma=\iota(z)$ is a density matrix, and the projections $P_i=\iota(p_i)$ on $\C^d$ with the state $X\mapsto\Tr(\sigma X)$ satisfy $\Tr(\sigma E_w)=\omega(\theta_W(w))$ for every word $w$, where $\theta_W(w)=U_{q_1}\cdots U_{q_m}e_W$ with $q_t\in\{p_i,e_W-p_i\}$. In particular, if $W$ is a unital subalgebra of a Euclidean Jordan algebra $V$ and $\omega(x)=\ip\rho x$ for a state $\rho$ of $V$, they form a sequential projective representation of the statistics of the tests from $\rho$: $\Tr(\sigma E_w)=\Pr_\rho(w)$. \end{lemma} \begin{proof} \emph{The element $z$.} The form $\tau(x,y)=\Tr(\iota(x)\iota(y))$ on $W$ is symmetric and bilinear, and positive definite: for $x\ne0$, $\iota(x)$ is a nonzero Hermitian matrix and $\Tr(\iota(x)^2)>0$. So the linear functional $\omega$ on the finite-dimensional space $W$ is $\tau(z,\cdot)$ for a unique $z$. \emph{$\sigma$ is a density matrix.} Write $z=\sum_k\mu_kc_k$ with distinct real $\mu_k$ and pairwise orthogonal nonzero idempotents $c_k$ of $W$ summing to $e_W$ (the spectral theorem in the Euclidean Jordan algebra $W$, Section~\ref{sec:set:eja}). Each $\iota(c_k)$ is a Hermitian idempotent, an orthogonal projection, and is nonzero, so $\Tr\iota(c_k)\ge1$. For $k\ne l$, $\iota(c_k)\iota(c_l)+\iota(c_l)\iota(c_k)=2\iota(c_k\jp c_l)=0$; for orthogonal projections $P,Q$ with $PQ+QP=0$, multiplying by $P$ on the left and on the right gives $PQ+PQP=0=PQP+QP$, so $PQ=QP$ and then $PQ=0$. Hence \[ \omega(c_k)=\tau(z,c_k)=\sum_l\mu_l\Tr\bigl(\iota(c_l)\iota(c_k)\bigr)=\mu_k\Tr\iota(c_k). \] The left side is $\omega(c_k^2)\ge0$. So every $\mu_k\ge0$ and $\sigma=\sum_k\mu_k\iota(c_k)\ge0$. Since $\iota$ is unital, $\Tr\sigma=\tau(z,e_W)=\omega(e_W)=1$. \emph{The words.} In an associative algebra $2X\jp(X\jp Y)-X^2\jp Y=XYX$, so a Jordan homomorphism satisfies $\iota(U_ab)=\iota(a)\iota(b)\iota(a)$. Unitality gives $\iota(e_W-p_i)=1-P_i$, so $\iota(q_t)=Q_t$ for every letter. Every $\theta_W(w)$ lies in $W$, and \[ \iota(\theta_W(w))=\iota(U_{q_1}\cdots U_{q_m}e_W)=Q_1\cdots Q_m\,1\,Q_m\cdots Q_1=E_w , \] using $Q_m^2=Q_m$. Therefore $\Tr(\sigma E_w)=\tau(z,\theta_W(w))=\omega(\theta_W(w))$. If $W$ is a unital subalgebra of $V$ and $\omega=\ip\rho\cdot$, then $\theta_W(w)=\theta(w)$ and $\omega(\theta(w))=\Pr_\rho(w)$. \end{proof} Apply the lemma to $W=B_{\mathrm{sp}}$, with unit $c_{\mathrm{sp}}=\pi_{\mathrm{sp}}(e)$, the idempotents $\pi_{\mathrm{sp}}(p_i)$, the embedding $\iota$ above and $\omega(x)=\ip\rho x$ for $x\in B_{\mathrm{sp}}$. Then $\omega(x^2)\ge0$, and $\omega(c_{\mathrm{sp}})=\ip\rho e=1$ because the discarded summands have weight zero. Since $\pi_{\mathrm{sp}}$ is a unital Jordan homomorphism, $\theta_{B_{\mathrm{sp}}}(w)=\theta_{\mathrm{sp}}(w)$. With $P_i=\iota(\pi_{\mathrm{sp}}(p_i))$ and the density matrix of the lemma, $\Tr(\sigma E_w)=\ip\rho{\theta_{\mathrm{sp}}(w)}=\Pr_\rho(w)$. Some $P_i$ may be $0$ or $1$. \begin{proof}[Proof of Theorem~\ref{thm:bridgeJ}] In (1), $B\neq\J$ implies that $B$ is special by the classification, and $B=\J$ is not special by Theorem~\ref{thm:tet:albert}. $(1)\Rightarrow(2)$: embed and construct the state. $(2)\Rightarrow(3)$ is trivial. $(3)\Rightarrow(1)$: if $B=\J$, the atoms are primitive in $B$, and a representation reproducing all words gives the unital homomorphism $\eta$ of Proposition~\ref{prop:bri:eta}, contradicting Theorem~\ref{thm:tet:albert}. The finite inequality is Corollary~\ref{cor:bri:finite}(a). \end{proof} \subsection{General Euclidean Jordan algebras}\label{sec:bri:general} Let $V$ be a Euclidean Jordan algebra, $p_1,\dots,p_n$ idempotents, $\rho$ a state, and $B=\Jord\langle e,p_1,\dots,p_n\rangle$. As a Euclidean Jordan algebra, \begin{equation}\label{eq:bri:decomp} B=B_{\mathrm{sp}}\oplus A_1\oplus\cdots\oplus A_m, \end{equation} where $B_{\mathrm{sp}}$ is the sum of the simple summands not isomorphic to $\J$ and each $A_k\cong\J$ (an \emph{Albert summand}), with unit $c_k$, a central idempotent of $B$. The state $\rho$ \emph{gives positive weight} to $A_k$ if $\ip\rho{c_k}>0$, and zero weight otherwise. \begin{definition}[$B$-primitive test]\label{def:bri:minimal} A test $p\notin\{0,e\}$ is \emph{$B$-primitive} if $p$ or $e-p$ is a primitive idempotent of $B$. \end{definition} This holds whenever $p$ or $e-p$ is primitive in $V$, so every test of a primitive idempotent of $V$ is $B$-primitive. It holds for every nontrivial idempotent when $V$ has rank at most three (e.g.\ $H_3(\mathbb K)$, $\J$, spin factors). It can fail when $V$ has rank at least four (Remark~\ref{rem:mixture}). \leannote{hand}{} \begin{theorem}[Sequential criterion for speciality]\label{thm:bridge} Let $V$ be a Euclidean Jordan algebra, $p_1,\dots,p_n\in V$ idempotents with generated algebra $B$ as in \eqref{eq:bri:decomp}, and $\rho$ a state of $V$. The model class is (I). \begin{enumerate}[label=(\alph*),leftmargin=2em] \item If $\rho$ gives zero weight to every Albert summand of $B$ (in particular, if $B$ is special), there is a sequential projective representation on $\C^d$ for some finite $d$, with a density-matrix state. \item Suppose every test $p_i\notin\{0,e\}$ is $B$-primitive. If $\rho$ gives positive weight to some Albert summand of $B$, there is no sequential projective representation: no Hilbert space over $\R$, $\C$ or $\Ha$, of any dimension, with projections of any rank and any state $\varphi$, reproduces all word probabilities. \end{enumerate} Thus, for $B$-primitive tests, a sequential projective representation exists if and only if $\ip\rho{c_k}=0$ for every Albert summand $A_k$ of $B$. \end{theorem} Part (a) is the construction of Section~\ref{sec:bri:converse}. Part (b) is the proof of Theorem~\ref{thm:bridgeJ} with three changes, carried out in Appendix~\ref{app:bri:general}: tests $p_i\in\{0,e\}$ are dropped and $p_i$ is replaced by $e-p_i$ where necessary, so that every test is primitive in $B$; Lemma~\ref{lem:purity} and the reconstruction then use only \eqref{eq:bri:lambda}; and injectivity is proved on one summand $A_k$ of positive weight, using $\ip\Omega{\eta(c_k)\Omega}=\ip\rho{c_k}>0$. Unlike Theorem~\ref{thm:bridgeJ}, part (b) depends on the state: when $B$ has an Albert summand and another summand, the same tests can have a representation from one state and none from another. \phantomsection\label{sec:bri:open}\begin{remark}[Mixtures of exceptional systems]\label{rem:mixture} It remains unresolved here whether $B$-primitivity can be dropped from Theorem~\ref{thm:bridge}(b). The simplest open case: in $V=\J\oplus\J$ take tests $p_i=(a_i,a_i')$ with $\{a_i\}$ and $\{a_i'\}$ atoms each generating $\J$ and the pairs generating $V$, and $\rho=\lambda\rho_1\oplus(1-\lambda)\rho_2$ with $0<\lambda<1$. Each $p_i$ has rank two in $B=V$ and is not $B$-primitive, so \eqref{eq:bri:lambda} and the purity relations fail. The table is a mixture of two tables that each have no representation (Theorem~\ref{thm:bridgeJ}), but the set of representable tables is convex, so this alone decides nothing. \end{remark} \subsection{Consequences}\label{sec:bri:cons} \leannote{hand}{} \begin{corollary}[Four tests decide; two or three never detect]\label{cor:four} For a Euclidean Jordan algebra $V$ and class (I), the following are equivalent: (i) $V$ is special; (ii) every finite family of tests of $V$ has a sequential projective representation from every state; (iii) every four tests of $V$ have a sequential projective representation from every state. In every $V$, any two and any three tests have a sequential projective representation on some $\C^d$ from every state. \end{corollary} \begin{proof} If $V$ is special, every generated algebra is special, so Theorem~\ref{thm:bridge}(a) gives (i)$\Rightarrow$(ii) (Appendix~\ref{app:bri:four}). If $V$ is not special it has a summand $A\cong\J$, with unit $c$. Four atoms generating $A$ (Theorem~\ref{thm:fourpoint}) generate $B=A\oplus\R(e-c)$, in which they are primitive, and the state $c/3$ gives $A$ positive weight, so Theorem~\ref{thm:bridge}(b) applies. Three tests always generate a special algebra, by Proposition~\ref{prop:three} (see also Section~\ref{sec:prior:generation}), so Theorem~\ref{thm:bridge}(a) applies (Appendix~\ref{app:bri:four}). \end{proof} A family need not be representable when each of its four-element subfamilies is (Theorem~\ref{thm:pentad}). Let $\mathcal Q(\mathbb K)$, respectively $\mathcal Q_{\mathrm{fin}} (\mathbb K)$, be the set of tables produced by sequential projective representations over $\mathbb K\in\{\R,\C,\Ha\}$ in arbitrary, respectively finite, dimension, and $\mathcal J(V)$ the set of tables produced by tests of $V$ from all its states. \provisionalref{Field classes}{prop:field}{Lean scope narrower}{(a) every inclusion, by constructions preserving all word probabilities and finite dimension: \leandecl{HilbertProjectiveData.complexExtensionModel_probability}, \leandecl{HilbertProjectiveData.quaternion}\allowbreak\leandecl{ExtensionModel_probability}, \leandecl{cyclicHilbertModel_probability}, \leandecl{HilbertProjectiveData.cyclicHilbert_finite}, \leandecl{Quaternionic}\allowbreak\leandecl{ProjectiveData.cyclicHilbert_finite_of_real}, \leandecl{complexification_finiteDimensional}, \leandecl{quaternionification_finiteDimensional}; the equalities are not stated as such. (b) and (c) are not formalized.} \begin{proposition}[Real, complex and quaternionic representations give the same tables] \label{prop:field} \leavevmode \begin{enumerate}[label=(\alph*),leftmargin=2em] \item $\mathcal Q_{\mathrm{fin}}(\R)=\mathcal Q_{\mathrm{fin}}(\C)=\mathcal Q_{\mathrm{fin}} (\Ha)$ and $\mathcal Q(\R)=\mathcal Q(\C)=\mathcal Q(\Ha)$. \item For every spin factor $V$, $\mathcal J(V)\subseteq\mathcal Q_{\mathrm{fin}}(\C)$, and the union of these sets over all spin factors is strictly smaller than $\mathcal Q_{\mathrm{fin}}(\C)$. \item $\mathcal J(V)\subseteq\mathcal Q(\C)$ iff $\mathcal J(V)\subseteq \mathcal Q_{\mathrm{fin}}(\C)$ iff $V$ is special. \end{enumerate} \end{proposition} \begin{proof} Part (a) follows from two standard constructions: the cyclic representation of Lemma~\ref{lem:bri:gns} turns every representation into a real one, finite-dimensional if the original is (Appendix~\ref{app:bri:gns}), and extension of scalars from $\R$ gives the converse. The rest follows from Theorem~\ref{thm:bridge}(a), Corollary~\ref{cor:four} and one explicit table (Appendix~\ref{app:bri:field}). \end{proof} \begin{remark}[The Albert algebra as a mathematical foil]\label{rem:bri:foil} Fix the number $n$ of tests. Write $\mathcal Q_n^{\rm fin}$ for the finite-dimensional quantum tables above, $\mathcal J_n(V)$ for the tables of $n$ ideal tests on $V$ from all its states, and $\mathcal J_n=\bigcup_V\mathcal J_n(V)$, with $V$ ranging over finite-dimensional Euclidean Jordan algebras. Then \begin{equation}\label{eq:bri:foil} \mathcal J_n=\operatorname{conv}\bigl(\mathcal Q_n^{\rm fin}\cup\mathcal J_n(\J)\bigr), \end{equation} where $\operatorname{conv}$ means finite convex combinations, without closure. Thus a linear inequality involving finitely many word probabilities, including the empty word if a constant term is needed, that holds on $\mathcal Q_n^{\rm fin}$ holds on all of $\mathcal J_n$ if and only if it holds on $\mathcal J_n(\J)$. Indeed, decompose $V=\bigoplus_k V_k$ into simple ideals with units $c_k$, and let $\omega(x)=\ip\rho x$ be the initial state functional. Put $\lambda_k=\omega(c_k)$. These weights are nonnegative and sum to one. For $\lambda_k>0$ the functional $\omega_k=\omega|_{V_k}/\lambda_k$ is positive and takes value one at $c_k$, hence is a state of $V_k$ in its own trace normalization. Every quadratic update acts componentwise, so the word table is $\sum_k\lambda_k f_k$, where $f_k$ is the table of the component idempotents from $\omega_k$; zero-weight terms vanish. By the classification (Section~\ref{sec:set:eja}), each $V_k$ is special or isomorphic to $\J$. In the special case Theorem~\ref{thm:bridge}(a) gives $f_k\in\mathcal Q_n^{\rm fin}$. Conversely, $\mathcal Q_n^{\rm fin}\subseteq\mathcal J_n$, since complex Hermitian matrices are Euclidean Jordan algebras, and direct sums of algebras, tests and weighted states realize every finite convex combination. This proves \eqref{eq:bri:foil}; linearity gives the inequality assertion. The unique \emph{simple} exceptional algebra supplies the only additional simple ingredient. Several Albert tables may be needed in the convex combination; neither their replacement by one Albert table nor a unique countermodel is asserted. The decomposition does not decide whether a particular mixture of Albert tables is quantum (Remark~\ref{rem:mixture}). \end{remark} \needspace{6\baselineskip} \provisionalref{Finite inequalities}{cor:bri:finite}{Lean scope narrower}{\leandecl{HilbertModelSeparation.uniformHilbertSeparator_generating_atoms}. Part (a) only.} \begin{corollary}[Finite state-independent inequalities]\label{cor:bri:finite} \leavevmode \begin{enumerate}[label=(\alph*),leftmargin=2em] \item In Theorem~\ref{thm:bridgeJ} with $B=\J$ there are $L$, a linear functional $\Lambda$ on the probabilities of words of length at most $L$ (the empty word included) and $\gamma>0$ such that $\Lambda\ge0$ on every sequential projective representation and $\Lambda\le-\gamma$ on the statistics from every state of $\J$. \item In Theorem~\ref{thm:bridge}(b), let $c_{\mathrm{ex}}=\sum_kc_k$. For each $\varepsilon>0$ the same holds for all states with $\ip\rho{c_{\mathrm{ex}}}\ge \varepsilon$, with $L,\Lambda,\gamma$ allowed to depend on $\varepsilon$. If $B$ has an Albert summand and $c_{\mathrm{ex}}\neq e$, no single choice of $L$, $\Lambda$ and $\gamma>0$ serves all states with $\ip\rho{c_{\mathrm{ex}}}>0$. \end{enumerate} \end{corollary} \begin{proof} The moment functionals of all representations form a compact convex set (Lemma~\ref{lem:bri:gns}, Tychonoff's theorem). If every length admitted a match, a nested intersection would give an exact all-word match, which Theorems~\ref{thm:bridgeJ} and~\ref{thm:bridge} exclude. Strict separation of disjoint compact convex sets in $\R^N$ then gives $\Lambda$ (Appendix~\ref{app:bri:compact}). \end{proof} Compactness gives no value of $L$, $\gamma$ or the coefficients; Section~\ref{sec:cer} gives the lengths seven and eleven for four atoms. \section{Three atoms fit a qutrit; four need not}\label{sec:four} \label{sec:fp} By Theorem~\ref{thm:bridgeJ}, the algebra generated by the tests decides whether a sequential projective representation reproduces them. One atom $p$ generates $\R p\oplus\R(e-p)$, and a qubit reproduces its test. Two atoms $p,q$ generate a subalgebra of $\spn\{e,p,q,p\jp q\}$, a copy of $\R\oplus H_2(\R)$ when they are distinct and not orthogonal, and a real qutrit with two rank-one projections at the right angle reproduces them. Conventions for octonions and for $\J$ are those of Appendix~\ref{app:conv}, and $\Ffour=\Aut(\J)$ is the automorphism group of $\J$, a compact Lie group of dimension $52$ (Appendix~\ref{app:A:F4}). A \emph{complex subalgebra} is a unital Jordan subalgebra of $\J$ isomorphic to $\HC$, and a \emph{quaternionic subalgebra} one isomorphic to $\HH$. The standard ones are $H_3(\C_i)$ and $H_3(\Ha)$, with $\C_i=\spn\{1,i\}$ and $\Ha=\spn\{1,i,j,ij\}$. \subsection{Three atoms always sit in a qutrit}\label{sec:fp:three} \provisionalref{Three atoms}{prop:three}{Same statement, different route}{\leandecl{complexTripleContract_holds}: the normalizing automorphisms are constructed in the development, without the facts on $\Ffour$ used in the printed proof. With only the space $\C^3$ recorded: \leandecl{AlbertChallenge.three_atoms}.} \begin{proposition}[Three atoms]\label{prop:three} Any three atoms $p_1,p_2,p_3$ of $\J$ lie in a common complex subalgebra $W\subseteq\J$. Consequently, for every state $\rho$ of $\J$, the statistics of the three tests from $\rho$ have a sequential projective representation (class (I)) on $\C^3$ with rank-one projections and a density-matrix state. \end{proposition} Earlier results on triangles and on algebras generated by three idempotents are recalled in Sections~\ref{sec:prior:invariants} and~\ref{sec:prior:generation}. The proof below also covers degenerate triples, and it gives rank-one projections, where a quaternionic subalgebra would give only rank-two projections on $\C^6$. \begin{proof} Lemma~\ref{lem:A:three} gives $g\in\Ffour$ with $gp_1,gp_2,gp_3\in H_3(\C_i)$; put $W=g^{-1}H_3(\C_i)$. The consequence is Lemma~\ref{lem:A:restrict}: a unital Jordan isomorphism $\psi:W\to H_3(\C)$ preserves the trace form, so it sends atoms to rank-one projections, and for the trace-orthogonal projection $\rho_W$ of $\rho$ onto $W$, $\psi(\rho_W)$ is a density matrix that reproduces every word in the three tests. \end{proof} \subsection{Thirteen traces}\label{sec:fp:traces} Four atoms modulo $\Ffour$ have $4\cdot16-52=12$ parameters: $\OP$ has dimension $16$, and when the four atoms generate $\J$ (which is the generic case, Theorem~\ref{thm:fourpoint}) an automorphism fixing them fixes every Jordan product of them and so is the identity. Traces of Jordan products serve as coordinates. \provisionalref{Thirteen traces}{def:traces}{Same statement}{\leandecl{fpOverlap}, \leandecl{fpTriangle}, \leandecl{fpDoublePair1}, \leandecl{fpOmega1} and their analogues.} \begin{definition}[The thirteen traces]\label{def:traces} For atoms $p_A,p_B,p_C,p_D$ of $\J$ define the six transition probabilities, the four shape invariants and the three double-pair traces \begin{gather*} o_{ij}=\ip{p_i}{p_j},\qquad R_{ijk}=\ip{p_i\jp p_j}{p_k},\\ t_1=\ip{p_A\jp p_B}{p_C\jp p_D},\quad t_2=\ip{p_A\jp p_C}{p_B\jp p_D},\quad t_3=\ip{p_A\jp p_D}{p_B\jp p_C}, \end{gather*} and the combinations $\omega_r=t_1+t_2+t_3-2t_r$ ($r=1,2,3$). \end{definition} The shape invariants are symmetric in their labels, because the trace form is associative, $\ip{x\jp y}z=\ip x{y\jp z}$. For atoms of a qutrit, $p_k=|a_k\rangle\langle a_k|$, one has $o_{ij}=|\langle a_i|a_j\rangle|^2$ and $R_{ijk}=\operatorname{Re}\langle a_i|a_j\rangle \langle a_j|a_k\rangle\langle a_k|a_i\rangle$, the real part of the Bargmann invariant, which Brehm calls the shape invariant of the triangle~\cite{Brehm1990}; so $R_{ijk}/\sqrt{o_{ij}o_{jk}o_{ki}}$ is the cosine of the Bargmann (geometric) phase. By Proposition~\ref{prop:three} the same holds for any three atoms of $\J$, computed in a complex subalgebra containing them. The double-pair traces carry the four-point data. The trace of a Jordan triple product against an atom is also useful: by associativity of the trace form, \begin{equation}\label{eq:fp:Q} \begin{aligned} Q_{j;klm}&:=\ip{p_j}{(p_k\jp p_l)\jp p_m+(p_m\jp p_l)\jp p_k-(p_k\jp p_m)\jp p_l}\\ &\phantom{:}=\ip{p_j\jp p_m}{p_k\jp p_l}+\ip{p_j\jp p_k}{p_m\jp p_l}-\ip{p_j\jp p_l}{p_k\jp p_m}, \end{aligned} \end{equation} so that $Q_{A;BCD}=\omega_2$, $Q_{A;CDB}=\omega_3$ and $Q_{A;DBC}=\omega_1$. \paragraph{Reading the traces from words.} Start in the maximally mixed state $e/3$, write words in chronological order with $i$ for ``yes'' on test $i$ and $\bar\imath$ for ``no''. The Jordan multiplication operator $L_p x=p\jp x$ of an idempotent is a signed sum of the identity and the two L\"uders updates, \begin{equation}\label{eq:fp:access} L_p=\tfrac12\bigl(\id+U_p-U_{e-p}\bigr), \end{equation} since both sides act as $1,\frac12,0$ on the Peirce spaces of $p$; the same identity holds for a projection $P$ acting on operators by $X\mapsto\frac12(PX+XP)$. So the operators generated by the L\"uders updates are those generated by the $L_{p_i}$, and right-normed products $p_{i_1}\jp(p_{i_2}\jp(\cdots))$ paired with a state are combinations of word probabilities. Since $U_pe=p$ and $U_px=\ip px\,p$ for an atom, a word that starts with $i$ and ends with $k$ has probability $\frac13\ip{p_i}{X p_k}$, where $X$ is the product of the intermediate $U$-operators, and \eqref{eq:fp:access} gives \begin{align*} o_{ij}&=3\Pr(i,j),\qquad R_{ijk}=\tfrac32\bigl[\Pr(i,k)+\Pr(i,j,k)-\Pr(i,\bar\jmath,k)\bigr],\\ t_1&=\tfrac34\textstyle\sum_{\sigma,\tau}\varepsilon_\sigma\varepsilon_\tau\, \Pr(A,\sigma_B,\tau_C,D), \end{align*} where $\sigma,\tau$ range over ``skip'', ``yes'', ``no'' with signs $+,+,-$; $t_2,t_3$ follow by relabelling. Transition probabilities need words of length two, shape invariants length three, double-pair traces length four. The thirteen traces are not independent. Because $\J$ has rank three, the fourth elementary symmetric function of any spectrum vanishes, and this gives one relation, linear in the shape invariants and double-pair traces. \provisionalref{Trace relation}{prop:fp:tracerel}{Lean scope narrower}{\leandecl{rankThree_fourIdempotent_trace_relation}, \leandecl{albert_fourAtom_trace_relation}. Stated for an explicit rank-three cubic model and for $\J$; general rank-three Euclidean Jordan algebras are not treated.} \begin{proposition}[The trace relation]\label{prop:fp:tracerel} Let $V$ be a Euclidean Jordan algebra of rank three in which primitive idempotents have trace one (for instance $\J$, or a complex or quaternionic subalgebra of $\J$). For any four primitive idempotents $p_A,p_B,p_C,p_D$ of $V$, \begin{equation}\label{eq:fp:tracerel} t_1+t_2+t_3=\tfrac12-\tfrac12\sum_{i0$, put $a=\sqrt{o_{AB}}$, $b=\sqrt{o_{AC}}$, $c=\sqrt{o_{AD}}$. Then \begin{equation}\label{eq:fp:Acal} \Acal=a^2b^2c^2\bigl(\det G-s^2\bigr), \end{equation} where the symmetric $3\times3$ matrix $G$ and the number $s$ are explicit rational functions of the traces and of $a,b,c$, displayed in \eqref{eq:A:G}. The polynomial is symmetric under all $24$ relabellings of $A,B,C,D$, invariant under $\Ffour$, and of degree three in each atom; expanded in the thirteen traces it has $52$ monomials (Appendix~\ref{app:A:Acal}). In part (iv) of the next theorem, $i$, $j$ and $\ell$ are the octonion units of the Cayley--Dickson construction $\Oct=\Ha\oplus\Ha\ell$ with $\Ha=\spn_\R\{1,i,j,ij\}$ (Appendix~\ref{app:A:oct}). \provisionalref{Four-point criterion}{thm:fourpoint}{Same statement}{\leandecl{thm_fourpoint}; the strict signs in (iv): \leandecl{lemma_A_normal_nondegenerate}. Parts (i) and (iii): \leandecl{AlbertChallenge.fourPoint_criterion}.} \begin{theorem}[Four-point criterion]\label{thm:fourpoint} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ and $\Acal$ the polynomial \eqref{eq:A:Acal52} in their thirteen traces. Then: \begin{enumerate}[label=(\roman*),leftmargin=2em] \item $\Acal\ge0$; \item $\Acal=0$ if and only if the four atoms lie in a common quaternionic subalgebra of $\J$; \item $\Acal>0$ if and only if $\Jord\langle e,p_A,p_B,p_C,p_D\rangle=\J$; \item if no two of the atoms coincide or are orthogonal, then there is $g\in\Ffour$ with $gp_i=v_iv_i^\dagger$ for the unit vectors \begin{equation}\label{eq:fp:normal} v_A=(1,0,0),\enspace v_B=(B_0,B_1,0),\enspace v_C=(C_0,\,C_1+C_2i,\,C_3),\enspace v_D=(D_0,X,Y), \end{equation} with $B_0,B_1,C_0,D_0>0$, $C_2,C_3\ge0$, $C_1\in\R$, $X=x_0+x_1i+x_2j$ and $Y=y_0+y_1i+y_2j+y_3\,ij+y_4\ell$ ($x_k,y_k$ real, $x_2,y_4\ge0$), and then \begin{equation}\label{eq:fp:assoc} \Acal=\tfrac14\,B_0^2B_1^4\,C_0^2C_2^2C_3^2\,D_0^2\,\bigl|[i,X,Y]\bigr|^2 =B_0^2B_1^4\,C_0^2C_2^2C_3^2\,D_0^2\,x_2^2y_4^2. \end{equation} Here $[x,y,z]=(xy)z-x(yz)$ is the associator. \end{enumerate} \end{theorem} That some four atoms generate $\J$ was known (Section~\ref{sec:prior:generation}). \begin{proof} \emph{Normal form.} Lemma~\ref{lem:A:normal} moves the atoms to \eqref{eq:fp:normal} using the stabiliser chain $\Ffour\supset\Spin(9)\supset\Spin(8)\supset\Spin(7)\supset G_2$; the first three vectors lie in $\C_i^3$, which is Proposition~\ref{prop:three} again. \emph{Traces become octonion data.} In the normal form put \[ g_1=B_0C_0+B_1z,\qquad g_2=B_0D_0+B_1X,\qquad g_3=C_0D_0+\bar zX+C_3Y,\qquad z=C_1+C_2i, \] the octonionic analogues of the three edge amplitudes not touching $A$, and $u_k=\operatorname{Im}g_k$. Lemma~\ref{lem:A:Aform} writes the thirteen traces in the $g_k$ \eqref{eq:fp:traceids}, without reassociating, and shows that $G$ is the Gram matrix $(\ip{u_k}{u_l})$ and $s=\ip{u_1\times u_3}{u_2}$, where $u\times v=\operatorname{Im}(uv)$. \emph{The decisive step.} For imaginary octonions the Harvey--Lawson identity \eqref{eq:A:HL} gives $\det G-s^2=\frac14|[u_1,u_3,u_2]|^2$. The associator is trilinear and alternating, and it vanishes when one argument is real or when the three arguments lie in a subalgebra generated by two octonions (Artin's theorem). Since $z,\bar z\in\C_i$, this leaves $[g_1,g_3,g_2]=B_1^2C_2C_3[i,Y,X]$, and \eqref{eq:fp:Acal} with $a=B_0$, $b=C_0$, $c=D_0$ gives the first form of \eqref{eq:fp:assoc}. For $X,Y$ as in \eqref{eq:fp:normal}, $[i,X,Y]=x_2[i,j,Y]=x_2y_4[i,j,\ell]$ and $|[i,j,\ell]|^2=4$, which is the second form. Lemma~\ref{lem:A:Aform} states this as a polynomial identity valid without normalising the vectors. \emph{(ii), ``if''.} Let $\psi:\HH\to\J$ be a unital Jordan monomorphism whose image is a quaternionic subalgebra containing the atoms. It maps Jordan frames to Jordan frames of $\J$, so it preserves the thirteen traces. Lemma~\ref{lem:A:zero}(a) brings the four atoms $\psi^{-1}(p_i)$ of the standard $H_3(\Ha)$ to the shape \eqref{eq:fp:normal} with $X,Y\in\Ha$, and Lemma~\ref{lem:A:Aform} gives $\Acal=\frac14(\cdots)|[i,X,Y]|^2=0$, since $\Ha$ is associative. \emph{(i).} If two atoms coincide or are orthogonal, they lie in a quaternionic subalgebra (Lemma~\ref{lem:A:degenerate}) and $\Acal=0$. Otherwise \eqref{eq:fp:assoc} applies. \emph{(ii), ``only if''.} Pairs that coincide or are orthogonal are covered by Lemma~\ref{lem:A:degenerate}. Otherwise $\Acal=0$ means $C_2C_3x_2y_4=0$ in \eqref{eq:fp:assoc}, and Lemma~\ref{lem:A:zero}(b) shows that in each of these four cases the atoms lie in a quaternionic subalgebra. \emph{(iii).} If $\Acal>0$, no pair coincides or is orthogonal, every factor in \eqref{eq:fp:assoc} is nonzero, and Lemma~\ref{lem:A:generate} shows that the atoms generate $\J$. Conversely, a quaternionic subalgebra is proper, so atoms that generate $\J$ have $\Acal\ne0$ by (ii), hence $\Acal>0$ by (i). \end{proof} \paragraph{The fourth atom in the complex subalgebra of the other three.} For atoms $p_1,p_2,p_3$ define the \emph{triangle factor} \begin{gather*} \Delta_W(p_1,p_2,p_3)=N(p_1+p_2+p_3)\,\bigl(o_{12}o_{23}o_{31}-R_{123}^2\bigr),\\ N(p_1+p_2+p_3)=1-o_{12}-o_{13}-o_{23}+2R_{123}. \end{gather*} For atoms of a qutrit, $p_k=|a_k\rangle\langle a_k|$, the first factor is $|\det(a_1,a_2,a_3)|^2$ and the second is $o_{12}o_{23}o_{31}\sin^2\phi$, with $\phi$ the Bargmann phase of the triangle; so $\Delta_W\ne0$ says that the three vectors are linearly independent and the triangle carries a non-real phase. \provisionalref{Window determinant}{prop:windowdet}{Same statement}{\leandecl{prop_windowdet}, \leandecl{prop_windowdet_zero}, for any complex subalgebra containing the three atoms; that it is the one they generate: \leandecl{lemma_A_window_generated}.} \begin{proposition}[The invariant as a determinant in a complex subalgebra]\label{prop:windowdet} Let $p_1,p_2,p_3$ be atoms of $\J$ with $\Delta_W(p_1,p_2,p_3)\ne0$. Then $W=\Jord\langle e,p_1,p_2,p_3\rangle$ is a complex subalgebra, and for every atom $p_4$ \[ \Acal(p_1,p_2,p_3,p_4)=\Delta_W(p_1,p_2,p_3)\cdot N(P_Wp_4), \] where $P_W$ is the trace-orthogonal projection of $\J$ onto $W$ and $N$ restricted to $W$ is the determinant of $3\times3$ Hermitian matrices. If $\Delta_W(p_1,p_2,p_3)=0$, then $\Acal(p_1,p_2,p_3,p_4)=0$ for every atom $p_4$. \end{proposition} \begin{proof} Assume first $\Delta_W\ne0$. By Lemma~\ref{lem:A:window}, after relabelling $p_1,p_2,p_3$ and applying some $g\in\Ffour$ the atoms $(p_1,p_2,p_3,p_4)$ take the normal form \eqref{eq:fp:normal} in the roles of $(A,B,C,D)$, with $gW=H_3(\C_i)$ and \begin{equation}\label{eq:fp:DeltaW} \Delta_W=B_0^2B_1^4\,C_0^2C_2^2C_3^2 . \end{equation} Since $g$ is orthogonal and preserves $N$, the computation can be made in the normal form. \emph{The fourth atom.} Write $X=X_c+X_\perp$, $Y=Y_c+Y_\perp$ with $X_c,Y_c\in\C_i$ and $X_\perp=x_2j$, $Y_\perp=y_2j+y_3\,ij+y_4\ell$ orthogonal to $\C_i$. The projection onto $H_3(\C_i)$ keeps the diagonal and replaces each off-diagonal entry by its $\C_i$-component. Since $\C_i\C_i^\perp+\C_i^\perp\C_i\subseteq\C_i^\perp$, the $23$ entry $X\bar Y$ of $p_4=v_Dv_D^\dagger$ projects to $X_c\bar Y_c+\kappa$, where $\kappa=x_2(y_2-y_3i)$ is the $\C_i$-component of $X_\perp\bar Y_\perp$. Thus \[ P_Wp_4=uu^\dagger+E,\qquad u=(D_0,X_c,Y_c),\qquad E=\begin{pmatrix}0&0&0\\0&x_2^2&\kappa\\0&\bar\kappa&|Y_\perp|^2\end{pmatrix}. \] Subtracting multiples of the first row, $\det(uu^\dagger+E)=D_0^2\bigl(x_2^2|Y_\perp|^2- |\kappa|^2\bigr)=D_0^2x_2^2y_4^2$. Multiplying by \eqref{eq:fp:DeltaW} gives the right side of \eqref{eq:fp:assoc}, which equals $\Acal$ by Lemma~\ref{lem:A:Aform}. If $\Delta_W=0$, Lemma~\ref{lem:A:window}(c) gives $\Acal=0$. \end{proof} \provisionalref{Rank trichotomy}{cor:fp:rank}{Same statement}{\leandecl{cor_fp_rank}.} \begin{corollary}[Rank trichotomy]\label{cor:fp:rank} In the situation of Proposition~\ref{prop:windowdet} with $\Delta_W\ne0$, $P_Wp_4$ is a state of the complex subalgebra $W$, and \begin{enumerate}[label=(\roman*),leftmargin=2em] \item $\operatorname{rank}P_Wp_4=1$ if and only if $p_4\in W$; \item $\operatorname{rank}P_Wp_4\le2$ if and only if $p_1,\dots,p_4$ lie in a common quaternionic subalgebra (which then contains $W$); \item $\operatorname{rank}P_Wp_4=3$ if and only if $p_1,\dots,p_4$ generate $\J$. \end{enumerate} Moreover $N(P_Wp_4)\le\frac1{27}$, with equality if and only if $P_Wp_4=e/3$. \end{corollary} \begin{proof} Positivity and trace one are Lemma~\ref{lem:A:restrict}. In the normal form above, $u$ has nonzero first entry and $E$ is supported on the last two coordinates, so $\operatorname{rank}P_Wp_4=1+\operatorname{rank}E$. Now $E=0$ iff $X,Y\in\C_i$ iff $p_4\in H_3(\C_i)$, which is (i). The lower block of $E$ has determinant $x_2^2y_4^2$, so $\operatorname{rank}E\le1$ iff $N(P_Wp_4)=0$ iff $\Acal=0$, which is (ii) by Theorem~\ref{thm:fourpoint}(ii); (iii) is the complement, by Theorem~\ref{thm:fourpoint}(iii). The last statement is the inequality between the arithmetic and geometric means of the eigenvalues of a positive matrix of trace one. \end{proof} The right-hand side of Proposition~\ref{prop:windowdet} is not manifestly symmetric; each choice of the fourth atom gives a different factorisation of the same number. The quaternionic rank bound also has a converse for complex states. Every complex density matrix of rank two is the complex part of a quaternionic pure state, as shown by Asorey, Scolarici and Solombrino~\cite[Prop.~4]{AsoreyScolariciSolombrino2007} and Masillo, Scolarici and Sozzo~\cite[Props.~2--3]{MasilloScolariciSozzo2009}. Indeed, write it as $xx^*+yy^*$ and take the quaternionic unit column $x+yj$, with $ji=-ij$. Its dyad $(x+yj)(x+yj)^*$ has complex part $xx^*+yy^*$. Taking $y=0$ covers rank one. \subsection{The tetrahedron identity} \label{sec:fp:tetra} Suppose a sequential projective representation, with projections $P_A,P_B,P_C,P_D$ and state $\varphi$, reproduces every word probability of the four tests of atoms $p_A,\dots,p_D$ from a state $\rho$. By Lemma~\ref{lem:bri:gns} the argument passes to the real cyclic representation: a real Hilbert space $\mathcal K$, orthogonal projections $P_i$ on it (with the same notation), and a unit vector $\Omega$ with $\Pr(w)=\|P_w\Omega\|^2$. By Lemma~\ref{lem:purity}, for every atom $p_a$ and every word $v$ the purity operator $P_aE_vP_a-c(v)P_a$ vanishes on $\mathcal K$, with $c(v)=\ip{p_a}{\theta(v)}$: in $\J$ a ``yes'' on an atom leaves a multiple of that atom. \provisionalref{Forced relations}{lem:fp:access}{Lean scope narrower}{\leandecl{triangleDefect_primitive_decomposition}, \leandecl{symmetricTripleDefect_primitive_decomposition}. Lean writes the differences of the two sides of (R2), (R3) as combinations of purity operators; the operator identities then follow from Lemma~\ref{lem:purity}, a step not stated as a declaration.} \begin{lemma}[Relations forced by the statistics]\label{lem:fp:access} In this situation, for all distinct labels, as operator identities on $\mathcal K$: \begin{align*} &\text{(R1)}\quad P_aP_bP_a=o_{ab}P_a,\\ &\text{(R2)}\quad P_a(P_bP_c+P_cP_b)P_a=2R_{abc}\,P_a,\\ &\text{(R3)}\quad P_a(P_bP_cP_d+P_dP_cP_b)P_a=2Q_{a;bcd}\,P_a . \end{align*} \end{lemma} \begin{proof} (R1) is the purity relation with $v=(b)$. For (R2) and (R3) use \eqref{eq:fp:access} for operators, $L_k(Y)=\frac12(P_kY+YP_k)=\frac12\bigl(Y+P_kYP_k-(1-P_k)Y(1-P_k)\bigr)$, and the identities, valid in any associative algebra, \[ P_kP_l+P_lP_k=2L_k(P_l),\qquad P_kP_lP_m+P_mP_lP_k=2\bigl[L_kL_l(P_m)+L_mL_k(P_l)-L_lL_k(P_m)\bigr]. \] Expanding each $L$ writes both left sides as signed combinations of sequential products $E_v$ with $v$ of length at most two, resp.\ three. The same formal combinations of L\"uders updates applied to $e$ in $\J$ give $2p_k\jp p_l$ and $2[p_k\jp(p_l\jp p_m)+p_m\jp(p_k\jp p_l)-p_l\jp(p_k\jp p_m)]$, by \eqref{eq:fp:access} in $\J$. Sandwiching by $P_a$ and applying purity term by term replaces each $P_aE_vP_a$ by $c(v)P_a$, and the sum of the $c(v)$ is the trace of the Jordan value against $p_a$, namely $2R_{akl}$, resp.\ $2Q_{a;klm}$ by \eqref{eq:fp:Q}. \end{proof} Two subspaces with projections $P,Q$ satisfying $PQP=\lambda P$ and $QPQ=\lambda Q$ are \emph{isoclinic}: all their principal angles are equal, with $\cos^2=\lambda$ (Wong~\cite{Wong1961}, Wolf~\cite{Wolf1963elliptic}, Hoggar~\cite[Prop.~1.2(b)]{Hoggar1977}); so (R1) is the isoclinic relation, and a representation of four atoms is a configuration of four pairwise isoclinic subspaces with parameters $o_{ab}$, subject to (R2) and (R3) (for algebras of projections with such relations see Section~\ref{sec:prior:isoclinic}). \paragraph{Holonomies at a vertex.} Fix a base vertex $j$ and order the other labels as $k,l,m$; assume all six transition probabilities are positive. The three \emph{triangle holonomies} at $j$ are \[ h_1=\frac{P_jP_kP_lP_j}{\nu_1},\quad h_2=\frac{P_jP_lP_mP_j}{\nu_2},\quad h_3=\frac{P_jP_mP_kP_j}{\nu_3}, \] \[ \nu_1=\sqrt{o_{jk}o_{kl}o_{lj}},\ \nu_2=\sqrt{o_{jl}o_{lm}o_{mj}},\ \nu_3=\sqrt{o_{jm}o_{mk}o_{kj}}, \] operators on the corner $P_j\mathcal K$, split as $h_a=H_a+N_a$ into Hermitian part $H_a=\frac12(h_a+h_a^*)$ and skew part $N_a=\frac12(h_a-h_a^*)$. On a qutrit with $P_i=|a_i\rangle\langle a_i|$, $h_1$ is multiplication by the Bargmann phase factor $\langle a_j|a_k\rangle\langle a_k|a_l\rangle\langle a_l|a_j\rangle/\nu_1$; in general it is the operator (non-Abelian) version, the holonomy of the geodesic triangle $jkl$. The shape invariants and double-pair traces are the real parts of the Bargmann invariants of three and four pure states~\cite{Bargmann1964}, whose relations for complex vectors are studied by Oszmaniec, Brod and Galv\~ao~\cite{OszmaniecBrodGalvao2021}. The relations attach to these operators the cosines of the Bargmann phases of the four triangles, \[ c_1=\frac{R_{jkl}}{\nu_1},\quad c_2=\frac{R_{jlm}}{\nu_2},\quad c_3=\frac{R_{jmk}}{\nu_3}, \quad c_4=\frac{R_{klm}}{\sqrt{o_{kl}o_{lm}o_{mk}}}, \] the three numbers \[ K_{12}=\frac{Q_{j;klm}}{\sqrt{o_{jk}o_{kl}o_{lm}o_{mj}}},\quad K_{23}=\frac{Q_{j;lmk}}{\sqrt{o_{jl}o_{lm}o_{mk}o_{kj}}},\quad K_{31}=\frac{Q_{j;mkl}}{\sqrt{o_{jm}o_{mk}o_{kl}o_{lj}}}, \] and the symmetric matrix $M$ with $M_{aa}=1-c_a^2$ and $M_{ab}=c_ac_b-K_{ab}$ ($a\ne b$). Put \begin{equation}\label{eq:fp:D} S=c_1c_2c_3-c_1M_{23}-c_2M_{13}-c_3M_{12},\qquad \Dtet=\det M-(c_4-S)^2 . \end{equation} This is the tetrahedron statistic, a function of the thirteen traces and hence of word probabilities of length at most four. \provisionalref{Tetrahedron identity}{thm:tetra}{Lean scope narrower}{\leandecl{tetrahedronEngine}. Lean proves the algebraic form, Theorem~\ref{thm:tetra:algebraic}; the passage to $\Dtet=0$ on a Hilbert space is not stated.} \begin{theorem}[Tetrahedron identity]\label{thm:tetra} Let $P_j,P_k,P_l,P_m$ be orthogonal projections on a real Hilbert space $\mathcal K$ of any dimension, with $P_j\ne0$. Let $o_{ab}>0$, $R_{abc}$ and $Q_{j;\cdot}$ be real numbers such that, as operator identities on $\mathcal K$, (R1) holds for all pairs, (R2) holds at the bases $j$ and $k$, and (R3) holds at the base $j$. Define $\Dtet$ from these numbers by \eqref{eq:fp:D}. Then $\Dtet=0$. \end{theorem} The precedent is Wolf's Lemma~7 in~\cite{Wolf1963elliptic}, discussed below. No division by $\sqrt{1-c_a^2}$ occurs, so real triangles ($c_a=\pm1$) and degenerate $M$ are covered. \begin{proof} The bounded operators on $\mathcal K$ form an associative real algebra in which the $P_a$ are idempotents satisfying the hypotheses of Theorem~\ref{thm:tetra:algebraic}. So $\Delta_jP_j=0$, and $\Delta_j=0$ because $P_j\ne0$. In \eqref{eq:tet:Dprime} the products of transition probabilities are $\kappa_a=\nu_a^2$, and with $D=\operatorname{diag}(\nu_1,\nu_2,\nu_3)$ the definitions give $\mathcal B=-DMD$ and $\sigma=\nu_1\nu_2\nu_3(c_4-S)$; for instance $b_{12}=\nu_1\nu_2(K_{12}-c_1c_2)$ because $\nu_1\nu_2=o_{jl}\sqrt{o_{jk}o_{kl}o_{lm}o_{mj}}$. Hence $0=\Delta_j=-\det\mathcal B-\sigma^2=(\nu_1\nu_2\nu_3)^2\,\Dtet$, and $\Dtet=0$ since every $o_{ab}>0$. \end{proof} On the corner $P_j\mathcal K$, with $1$ denoting $P_j$, the steps of the proof of Theorem~\ref{thm:tetra:algebraic}, divided by the $\nu_a$, are statements about the holonomies. Step~(b) says that each $h_a$ is unitary and step~(a) that its Hermitian part is $H_a=c_a$, so $N_a^2=-M_{aa}$, and step~(d) gives the Clifford relations \begin{equation}\label{eq:fp:clifford} N_aN_b+N_bN_a=-2M_{ab}\qquad(a,b=1,2,3). \end{equation} After the contractions $P_lP_jP_l=o_{jl}P_l$ and $P_mP_jP_m=o_{jm}P_m$ the product $h_1h_2h_3$ is a multiple of $P_jP_kP_lP_mP_kP_j$, the holonomy of the far face, and step~(e) says \begin{equation}\label{eq:fp:face} \tfrac12\bigl(h_1h_2h_3+(h_1h_2h_3)^*\bigr)=c_4 . \end{equation} Expanding $h_1h_2h_3=(c_1+N_1)(c_2+N_2)(c_3+N_3)$, the left side is $S+\Gamma$ with the volume element $\Gamma=\frac12(N_1N_2N_3-N_3N_2N_1)$ of the Clifford relations, and step~(f), the one place where associativity is used, is \begin{equation}\label{eq:fp:volume} \Gamma=N_1N_2N_3+M_{23}N_1-M_{13}N_2+M_{12}N_3,\qquad \Gamma^2=\det M \end{equation} (Lemma~\ref{lem:A:clifford}). So $\Gamma=c_4-S$ and $(c_4-S)^2=\det M$. \begin{figure}[htbp] \centering \begin{tikzpicture}[scale=1.0,every node/.style={font=\small}] \coordinate (j) at (0,0); \coordinate (k) at (90:1.9); \coordinate (l) at (210:1.9); \coordinate (m) at (330:1.9); \draw[thick] (k)--(l)--(m)--cycle; \draw (j)--(k) (j)--(l) (j)--(m); \fill (j) circle (1.5pt) node[above right=1pt] {$j$}; \fill (k) circle (1.5pt) node[above] {$k$}; \fill (l) circle (1.5pt) node[left] {$l$}; \fill (m) circle (1.5pt) node[right] {$m$}; \node at (150:0.75) {$h_1$}; \node at (270:0.62) {$h_2$}; \node at (30:0.75) {$h_3$}; \node[align=left,anchor=west] at (2.4,0.2) {far face $klm$:\\ cosine $c_4$}; \end{tikzpicture} \caption{Schematic, drawn flat with the vertex $j$ in the middle. The holonomies $h_1,h_2,h_3$ go around the triangles $jkl$, $jlm$, $jmk$; their product, after two contractions by (R1), is the holonomy around the far face $klm$. In a sequential projective representation this closes up ($\Dtet=0$); for four atoms of $\J$ in general position it fails by a quarter of a squared associator.} \label{fig:fp:tetra} \end{figure} The argument never treats $\Gamma$ as a sign $\pm1$, chooses no Clifford module and does not normalise $M$, so a direct sum of models with different Clifford sectors gives no escape. It needs the relations only on finitely many vectors (Section~\ref{sec:cer:eleven}). \provisionalref{Dtet and A}{prop:Dacal}{Same statement}{\leandecl{prop_Dacal_poly}; \leandecl{tetrahedronDiscriminant_eq_overlap_mul_fourPointInvariant} is a variant normalised at one vertex.} \begin{proposition}[The tetrahedron statistic is the invariant]\label{prop:Dacal} For every choice of base $j$ and order $k,l,m$, \[ \Dtet\cdot o_{AB}\,o_{AC}\,o_{AD}\,o_{BC}\,o_{BD}\,o_{CD}=\Acal , \] as an identity of rational functions in the thirteen traces regarded as independent variables with positive transition probabilities. In particular $\Dtet$ does not depend on the base or the order. \end{proposition} \begin{proof} By the symmetry of $\Acal$ it suffices to take $j=A$, $(k,l,m)=(B,C,D)$. Put $d=\sqrt{o_{BC}}$, $e'=\sqrt{o_{BD}}$, $f=\sqrt{o_{CD}}$, let $G,s$ be as in \eqref{eq:fp:Acal}, and let $G'$ be $G$ with indices in the order $(1,3,2)$ (edges $BC$, $CD$, $BD$, matching $h_1,h_2,h_3$). Substituting the definitions gives \[ M=T\,G'\,T,\qquad T=\operatorname{diag}\Bigl(\frac1d,\frac1f,-\frac1{e'}\Bigr),\qquad c_4-S=\frac{s}{d\,e'f}. \] Hence $\Dtet=(\det G-s^2)/(de'f)^2$; multiplying by $\prod o_{ij}=a^2b^2c^2d^2e'^2f^2$ and using \eqref{eq:fp:Acal} gives $\Acal$. No relation among the traces is used, in particular not \eqref{eq:fp:tracerel}. \end{proof} \provisionalref{General obstruction}{cor:fp:noideal}{Same statement}{\leandecl{LengthSeven.Uniform.noideal_algebraic}, \leandecl{LengthSeven.Uniform.noideal_hilbert}, \leandecl{LengthSeven.Uniform.noideal_rightQuaternionic}, from every state; also \leandecl{FiveTest.cor_fp_noideal}.} \begin{corollary}[No sequential projective representation]\label{cor:fp:noideal} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$. The model class is (I). No sequential projective representation reproduces all word probabilities of the four tests from $e/3$. \end{corollary} \begin{proof} If $\Acal>0$, no two atoms coincide or are orthogonal (Theorem~\ref{thm:fourpoint}), so all $o_{ij}\in(0,1)$. In a reproducing representation Lemma~\ref{lem:fp:access} supplies the hypotheses of Theorem~\ref{thm:tetra}, with $P_j\ne0$ on $\mathcal K$ because $\|P_j\Omega\|^2=\Pr(j)=\frac13$. So $\Dtet=0$, and Proposition~\ref{prop:Dacal} gives $\Acal=\Dtet\prod o_{ij}=0$, a contradiction. \end{proof} \leannote{formal}{$\Acal=0$: \leandecl{fourPoint_zero_c6_representation}, with fixed rank-two projections; with only the space $\C^6$ recorded, \leandecl{AlbertChallenge.zero_invariant_model}.} This is the four-atom case of Theorem~\ref{thm:bridgeJ}, which covers every state, by a different route. If $\Acal=0$, the atoms lie in a quaternionic subalgebra and the standard embedding $\HH\to H_6(\C)$ gives a representation on $\C^6$ with rank-two projections. \paragraph{Octonionic holonomies.} In the normal form \eqref{eq:fp:normal} put $h_1=g_1/|g_1|$, $h_2=g_3/|g_3|$, $h_3=\bar g_2/|g_2|$: three unit octonions, the phases of the triangles $ABC$, $ACD$, $ADB$ read with all edges at $A$ real and positive. By \eqref{eq:fp:traceids}, $c_a=\operatorname{Re}h_a$, $M$ is the Gram matrix of $v_a=\operatorname{Im}h_a$, and $c_4=\operatorname{Re}((h_1h_2)h_3)$. Imaginary octonions satisfy the Clifford relations $v_av_b+v_bv_a=-2\ip{v_a}{v_b}$, like the $N_a$, so the octonionic phases obey everything a sequential projective representation demands of the holonomies, including the scalar relation \eqref{eq:fp:face}. In any associative model these relations fix $\Gamma^2=\det M$ by regrouping products. For octonions, $\operatorname{Re}((h_1h_2)h_3)-S=\operatorname{Re}((v_1v_2)v_3)=-\ip{v_1\times v_2}{v_3}$, and the Harvey--Lawson identity gives \[ \Dtet=\det M-\ip{v_1\times v_2}{v_3}^2=\tfrac14\bigl|[v_1,v_2,v_3]\bigr|^2 =\tfrac14\bigl|[h_1,h_2,h_3]\bigr|^2 : \] the octonions miss by a squared associator. The tetrahedron identity holds exactly when the three phase axes lie in an associative $3$-plane, the imaginary part of a quaternion subalgebra. The same skeleton appears in Wolf's study of totally geodesic families of pairwise isoclinic subspaces~\cite[Prop.~1, Lemma~7, Thm.~2]{Wolf1963elliptic}. When the projective lines of such a family have real dimension four, the family supplies three anticommuting unitaries satisfying the Hurwitz equations, and its closure under the Lie triple product forces $A_1A_2A_3=\pm1$ (pp.~454--455). The Hurwitz equations alone do not force this, since the sum of the two irreducible Clifford modules with opposite volume elements satisfies them with $A_1A_2A_3$ not scalar. The same argument rules out lines of dimension greater than four, which excludes the Cayley plane. Wolf assumes a whole totally geodesic family in a finite-dimensional Grassmannian over an associative field; here the input is four points in an arbitrary representation, the parameters $o_{ab}$ vary, $M$ may be degenerate, and the failure is measured by $\Acal$. Neither result implies the other. \subsection{How large can the invariant be?}\label{sec:fp:max} \provisionalref{Maximum}{prop:max}{Same statement}{\leandecl{prop_max}, \leandecl{maximizerConfig_traces}, \leandecl{fourPointInvariant_maximizerConfig}. The bound and the value at the maximiser: \leandecl{AlbertChallenge.quadrupleM_spec}.} \begin{proposition}[Maximum]\label{prop:max} For all atoms $p_A,p_B,p_C,p_D$ of $\J$, $0\le\Acal\le\frac4{19683}=\frac4{3^9}$. Equality holds on exactly one $\Ffour$-orbit of ordered quadruples, represented by the atoms $v_iv_i^\dagger$ with \begin{equation}\label{eq:fp:maximiser} v_A=(1,0,0),\quad v_B=\tfrac1{\sqrt3}(1,\sqrt2,0),\quad v_C=\tfrac1{\sqrt3}(1,i,1),\quad v_D=\tfrac1{\sqrt3}(1,j,\ell). \end{equation} There all transition probabilities are $\frac13$, all shape invariants $\frac19$, all double-pair traces $\frac1{27}$, and $\Dtet=\frac4{27}$. \end{proposition} \begin{proof} Configurations with a coinciding or orthogonal pair have $\Acal=0$. Otherwise, in the normal form, $\Acal=(B_0^2B_1^4)(C_0^2C_2^2C_3^2)(D_0^2x_2^2y_4^2)$ with unit vectors. The first factor is $u(1-u)^2$ with $u=B_0^2$, at most $\frac4{27}$, with equality only at $u=\frac13$. The second and third are products of three nonnegative numbers with sum at most one, so at most $\frac1{27}$, with equality only when the three numbers are $\frac13$ and all other coordinates vanish. This gives the bound, and equality forces the coordinates of \eqref{eq:fp:maximiser}; since every maximiser has no coinciding or orthogonal pair, the normal form applies to it, so the maximisers form one orbit. The traces are a direct computation, and $\Dtet=\Acal/\prod o_{ij}=\frac4{3^9}\cdot3^6$. \end{proof} Roelands and Tiersma~\cite[Lemma~6.5]{RoelandsTiersma2026} show by direct computation that a closely related quadruple of atoms generates $\J$. \provisionalref{Triangle bound}{lem:fp:DeltaWmax}{Same statement}{\leandecl{lem_fp_DeltaWmax}.} \begin{lemma}[Bound for the triangle factor]\label{lem:fp:DeltaWmax} For any three atoms of $\J$, $\Delta_W(p_1,p_2,p_3)\le\frac4{729}$, with equality if and only if $o_{12}=o_{13}=o_{23}=\frac13$ and $R_{123}=\frac19$. \end{lemma} \begin{proof} Let $\Delta_W\ne0$. Lemma~\ref{lem:A:window}(b), applied with $p_4=p_1$, gives $00$ says that the atoms generate $\J$ (Theorem~\ref{thm:fourpoint}). For a table of word probabilities, the question is whether some sequential projective representation reproduces it. The thirteen traces are read from words of length at most four, so a representation has them too, and with them a value of $\Dtet$. But a representation can reproduce all thirteen traces of the generating quadruple $\mathsf S$ (Section~\ref{sec:cer}, paragraph ``Two configurations''), with $\Dtet=\frac1{12}>0$ (Proposition~\ref{prop:Dfails}), and \companionref{section ``The left-multiplication model''} gives a real representation that reproduces all thirteen traces of any generating quadruple. The tetrahedron identity forces $\Dtet=0$ only through the purity relations, which are statements about longer words. Corollary~\ref{cor:fp:noideal} uses them on all words. \section{Finite certificates}\label{sec:cer} Four tests of atoms with $\Acal>0$ have no sequential projective representation that matches every word (Section~\ref{sec:fp}). Finitely many words already suffice, and from $e/3$ the least length is seven. Throughout, the model class is class (I) of Definition~\ref{def:idealmodel}. Words are written in chronological order with $i$ for ``yes'' and $\bar\imath$ for ``no'' on test $i$; a word is \emph{reduced} if no two consecutive letters belong to the same test. In every sequential projective representation $Q_aQ_a=Q_a$ and $P_i(1-P_i)=0$, so the probability of any word is that of a reduced word or zero; only reduced words are therefore listed. The notation below uses $E_w$, the word operators $P_w$ and the real cyclic representation $(\pi,\Omega)$ of Lemma~\ref{lem:bri:gns}, the purity operators $z_q(v)$ of Lemma~\ref{lem:purity}, the relations (R1)--(R3) of Lemma~\ref{lem:fp:access}, and the holonomies $h_a=H_a+N_a$, the numbers $c_a$, $S$, the matrix $M$ and the volume element $\Gamma=\frac12(N_1N_2N_3-N_3N_2N_1)$ of Section~\ref{sec:fp:tetra}. For four atoms with $\Acal>0$, let the \emph{minimum separating word length} $L_{\min}$ be the least $L$ such that no sequential projective representation reproduces the octonionic probabilities from $e/3$ of all words of length at most $L$. \provisionalref{Uniform length seven}{thm:cer:uniform}{Lean scope narrower}{(a) \leandecl{AlbertChallenge.seven_steps_exclude}, \leandecl{LengthSeven.Uniform.uniform_lengthSeven_e3_algebraic}, \leandecl{LengthSeven.Uniform.uniform_lengthSix_pA_algebraic} and their Hilbert and quaternionic versions; $L_{\min}=7$: \leandecl{AlbertChallenge.least_separating_length}, \leandecl{MatchingModels.lMin_eq_seven}. The minimum is taken over Hilbert spaces in Lean's lowest universe; the exclusions hold in every universe. (b) \leandecl{Ends.uniform_b}; also a margin from $p_A$ at length six, \leandecl{Ends.uniform_b_pA}. Below Lemma~\ref{lem:cer:short} the formal proof of (a) fixes the Gram table directly and does not use Lemma~\ref{lem:cer:reduce} (Appendix~\ref{app:lean-plan}).} \begin{theorem}[Words of length seven suffice for every generating quadruple]\label{thm:cer:uniform} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$; the model class is (I). \begin{enumerate}[label=(\alph*),leftmargin=2em] \item No sequential projective representation, over $\R$, $\C$ or $\Ha$, of any dimension, with projections of any rank and any state, reproduces from $e/3$ the probability of every word of length at most seven that begins with ``yes'' on $A$; none reproduces from $p_A$ the probability of every word of length at most six. Hence $L_{\min}=7$. \item There is $\delta>0$, depending on the quadruple, such that every sequential projective representation differs from the octonionic table from $e/3$ by at least $\delta$ on some word of length at most seven that begins with ``yes'' on $A$. \end{enumerate} \end{theorem} The proof is given in Section~\ref{sec:cer:horizon}, after the lemmas it uses. The lower bound $L_{\min}\ge7$ is the next proposition, where $\iota_{\Ha}:\HH\to H_6(\C)$ denotes the standard embedding. \provisionalref{Six-step models}{prop:cer:universalsix}{Lean scope narrower}{\leandecl{MatchingModels.prop_cer_universalsix_witness}, \leandecl{MatchingModels.prop_cer_universalsix_c12}, \leandecl{AlbertChallenge.six_step_model}; the last sentence: \leandecl{AlbertChallenge.least_separating_length}, \leandecl{MatchingModels.lMin_eq_seven}, \leandecl{MatchingModels.no_classI_linearEq_le_six}, and for inequalities \leandecl{MatchingModels.no_classI_linearIneq_le_six}. The minimum is taken over Hilbert spaces in Lean's lowest universe; the exclusions hold in every universe.} \begin{proposition}[Six-step models for every generating quadruple]\label{prop:cer:universalsix} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$. There are atoms $p_i^+$ and $p_i^-$ ($i=A,\dots,D$) of $\HH$ and $\lambda\in(0,1)$ such that the projections $\iota_{\Ha}(p_i^+)\oplus\iota_{\Ha}(p_i^-)$ on $\C^{12}$ with the density matrix $\lambda I_6/6\oplus(1-\lambda)I_6/6$, a mixture of two quaternionic quadruples, form a sequential projective representation that reproduces from $e/3$ every word of length at most six. Consequently, with Theorem~\ref{thm:cer:uniform}, $L_{\min}=7$ for every generating quadruple, and no linear equality valid in every sequential projective representation and supported on words of length at most six is violated by its table from $e/3$. \end{proposition} \begin{proof} Appendix~\ref{app:cert:universalsix} constructs the two quaternionic quadruples, from the two ends of a chord of complex density matrices through the projection of the fourth atom onto the complex subalgebra of the other three. They have the transition probabilities and shape invariants of the Albert quadruple, and the same average double-pair traces. It proves that every word probability through length six is affine in those double-pair traces when the other data are fixed. The mixture therefore reproduces these words, so $L_{\min}\ge7$; the upper bound is Theorem~\ref{thm:cer:uniform}. The equality assertion follows because the matching model is itself a sequential projective representation. \end{proof} \paragraph{Why six steps are affine.} Expand a word probability from $e/3$ with $U_px=\ip px\,p$ for a ``yes'' and $U_{e-p}=\id-2L_p+U_p$ for a ``no''. Each ``yes'' splits a term into trace factors, and the intervals of positions used by two factors share at most the splitting position. A factor that involves all four atoms within at most seven positions contains some atom only once; moving that atom to the outside of the trace pairing, and using the complex subalgebra generated by the other three, shows that the factor is affine in the double-pair traces. Two such factors would need at least $4+4-1=7$ positions. So through length six each term has at most one factor that depends on the double-pair traces, and it depends on them affinely (Appendix~\ref{app:cert:universalsix}). The same theorem gives seven steps from every preparation, provided the first outcome may be a ``yes'' on any of the four tests. \provisionalref{Every preparation}{cor:cer:everystate}{Same statement}{\leandecl{BlockB.cor_cer_everystate}, \leandecl{BlockB.everystate_lengthSeven_algebraic} and its Hilbert and quaternionic versions.} \begin{corollary}[Seven steps from every state]\label{cor:cer:everystate} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$, and let $\rho$ be any state of $\J$; the model class is (I). No sequential projective representation, over $\R$, $\C$ or $\Ha$, of any dimension, with projections of any rank and any state, reproduces from $\rho$ the probability of every word of length at most seven whose first outcome is ``yes''. For $\rho=e/3$ the length seven cannot be replaced by six. \end{corollary} \begin{proof} Let $y=p_A+p_B+p_C+p_D$. If $y$ were not invertible, an atom $r$ with $\ip ry=0$ would have $\ip r{p_i}=0$ for each $i$, so every $p_i$ would lie in $\J_0(r)$ (Lemma~\ref{lem:set:support}) and the atoms would generate a subalgebra of the proper subalgebra $\R r\oplus\J_0(r)$, against Theorem~\ref{thm:fourpoint}(iii). So $y$ is invertible and positive, $\ip\rho y>0$, and $a=\ip\rho{p_j}>0$ for some $j$. Suppose a representation with state $\varphi$ reproduces the words. Then $\varphi(P_j)=a$, and $\varphi_j(X)=a^{-1}\varphi(P_jXP_j)$ is again a state. For a word $v$ of length at most six, $E_{(j,v)}=P_jE_vP_j$, and in $\J$ the ``yes'' on $j$ leaves $U_{p_j}\rho=a\,p_j$; hence $\varphi_j(E_v)=a^{-1}\Pr_\rho(j,v)=\Pr_{p_j}(v)$. So the same projections with the state $\varphi_j$ reproduce from $p_j$ every word of length at most six, which Theorem~\ref{thm:cer:uniform}(a), with the labels exchanged so that $j$ plays the part of $A$, forbids. The last sentence is Proposition~\ref{prop:cer:universalsix}. \end{proof} From the maximally mixed state the first outcome may be fixed to be a ``yes'' on $A$; from a general state no single first test will do, since $\rho$ may give $p_A$ weight zero. \leannote{narrower}{$\mathsf S$: \leandecl{AlbertChallenge.quadrupleS_spec}, that its elements are atoms and $\Acal=\frac1{5832}$; its traces: \leandecl{BlockB.isTracesS_fixedS}. $\mathsf M$: \leandecl{AlbertChallenge.quadrupleM_spec}, \leandecl{BlockB.isTracesM_maximizerConfig}. The value of $\Dtet$ for $\mathsf S$ is not stated in this form.} \paragraph{Two configurations.} Let $(i,j,\ell)$ be a basic triple of unit imaginary octonions ($i\perp j$, and $\ell\perp1,i,j,ij$; Appendix~\ref{app:A:F4}). The \emph{maximal configuration} $\mathsf M$ is the maximiser \eqref{eq:fp:maximiser} of Proposition~\ref{prop:max}, the atoms $vv^\dagger/|v|^2$ of \[ v_A=(1,0,0),\quad v_B=(1,\sqrt2,0),\quad v_C=(1,i,1),\quad v_D=(1,j,\ell). \] The configuration $\mathsf S$ replaces $v_B$ by $(1,1,0)$; all its vectors have rational octonion coordinates, so its traces are rational. Both generate $\J$. For $\mathsf M$ all transition probabilities are $\frac13$, all shape invariants $\frac19$, all double-pair traces $\frac1{27}$, $\Dtet=\frac4{27}$ and $\Acal=\frac4{19683}$. For $\mathsf S$, $o_{AB}=\frac12$ and the other transition probabilities are $\frac13$, $(R_{ABC},R_{ABD},R_{ACD},R_{BCD})=(\frac16,\frac16,\frac19,\frac19)$, $t_1=t_2=t_3=\frac1{18}$, $\Dtet=\frac1{12}$ and $\Acal=\frac1{5832}$. The traces of both are derived in the proof of Proposition~\ref{prop:horizon}, and the values of $\Acal$ follow from \eqref{eq:A:Acal52}: for $\mathsf S$, with $\omega_r=\frac1{18}$, its eight groups of terms are $\frac1{486}$, $-\frac1{486}-\frac1{729}$, $-\frac1{729}$, $-\frac7{5832}$, $\frac5{2916}$, $\frac1{729}$, $\frac1{729}$ and $-\frac1{2916}$, with sum $\frac1{5832}$, and $\Dtet=\Acal/\prod o_{ab}=486/5832$ (Proposition~\ref{prop:Dacal}); for $\mathsf M$ see Proposition~\ref{prop:max}. Appendix~\ref{app:cert:conv} explains why the configurations, and hence every number below, do not depend on the choice of octonion basis. \subsection{Finite lists}\label{sec:cer:frame} \paragraph{The model side at length $L$.} The following lemma is the finite-level form of the moment-matrix characterisation of Budroni, Moroder, Kleinmann and G\"uhne~\cite[eqs.~(7)--(8) and App.~A3]{BMKG2013}. Their appendix sketches the converse by reference to the proof of Theorem~8 of Navascu\'es, Pironio and Ac\'{\i}n~\cite{NPA2008}, which treats the limit of all orders; the truncated statement needs only the short argument below. Let $\mathcal W_L$ be the set of reduced words of length at most $L$ in the letters $i,\bar\imath$ ($i=1,\dots,n$), the empty word included. A \emph{table on $\mathcal W_L$} is a function $f:\mathcal W_L\to[0,1]$ with $f(\varnothing)=1$. \provisionalref{Finite reconstruction}{lem:cer:recon}{Same statement, different route}{\leandecl{finiteGram_iff_finiteRealModel}, \leandecl{AlgebraicProjectiveData.finiteGram}: necessity is proved from the positive functional of an algebraic model, not through the cyclic representation.} \begin{lemma}[Reconstruction at finite length]\label{lem:cer:recon} A table $f$ on $\mathcal W_L$ is the restriction of the table of a sequential projective representation if and only if there is a real positive semidefinite matrix $\mathcal G$, indexed by $\mathcal W_L$, with \begin{enumerate}[label=(M\arabic*),leftmargin=2.8em] \item $\mathcal G_{w,w}=f(w)$ for every $w\in\mathcal W_L$; \item $\mathcal G_{u,vi}+\mathcal G_{u,v\bar\imath}=\mathcal G_{u,v}$ for all $u\in\mathcal W_L$ and all $v\in\mathcal W_{L-1}$ whose last letter is not a letter of test $i$; \item $\mathcal G_{u,v}=0$ whenever $u$ and $v$ end with the two different outcomes of the same test. \end{enumerate} In that case there is a representation on a real Hilbert space of dimension $\operatorname{rank} \mathcal G$ with a vector state. \end{lemma} The proof is in Appendix~\ref{app:cert:recon}. The set of tables on $\mathcal W_L$ of sequential projective representations is therefore convex and compact: direct sums give convex combinations, and $|\mathcal G_{u,v}|\le1$. \paragraph{The group algebra.} With $\Sigma_i=2P_i-1$ as in Section~\ref{sec:bri:recon}, in every representation $\Sigma_i$ is a self-adjoint unitary. Let $\R[\mathbb Z_2^{*n}]$ be the real group algebra of the free product $\mathbb Z_2^{*n}=\mathbb Z_2*\cdots*\mathbb Z_2$ ($n$ factors) with generators $r_1,\dots,r_n$; its basis is the set of reduced words in the $r_i$, and $u\mapsto u^{\mathrm{rev}}$ (reversal) is its involution. A formal word combination $\Lambda=\sum_w\lambda_w[w]$ has the element $E(\Lambda)=\sum_w\lambda_w E_w$ of $\R[\mathbb Z_2^{*n}]$, with each $P_i$ written as $\frac12(1+r_i)$. On the octonionic side, $\Lambda$ has the effect $\theta(\Lambda)=\sum_w\lambda_w\theta(w)\in\J$ of Section~\ref{sec:bri:null}; no map from $\R[\mathbb Z_2^{*n}]$ to $\J$ is needed. \subsection{Words of length seven suffice and length six does not}\label{sec:cer:horizon} The proof of Theorem~\ref{thm:cer:uniform} builds, from a representation that matches the words, three real vectors whose Gram matrix the word probabilities force to have negative determinant. It runs in four steps, which this subsection takes in order. First, conditioning on ``yes'' to $A$ gives hypothesis (H) below, and the words force relations on short histories (Lemma~\ref{lem:cer:short}). Second, these relations determine every walk value of length at most eight (an inner product $\ip\xi{P_{w_1}\cdots P_{w_n}\xi}$, defined below) as a fixed polynomial in the traces (Lemma~\ref{lem:cer:reduce}). Third, an explicit indefinite form, the comparison form, satisfies the same relations (Lemma~\ref{lem:cer:tablemodel}), so the words fix the Gram matrix of twenty-eight short path vectors to be their Gram matrix under that form (Lemma~\ref{lem:cer:gram28}). Fourth, in that form three combinations of them have Gram determinant $-4\hat D^3/\prod_x(1-o_{Ax})$, where $\hat D=\Acal/(o_{AB}o_{AC}o_{AD})^3$ (Lemma~\ref{lem:cer:negative}); the proof of the theorem follows that lemma. The rest of the subsection treats $\mathsf M$ and $\mathsf S$ by one explicit vector (Theorem~\ref{thm:cer:seven}), shows that six steps do not suffice for them (Proposition~\ref{prop:horizon}), and explains why the six-step mixture fails at seven (Remark~\ref{rem:cer:variance}, Proposition~\ref{prop:cer:schedule}). On a first reading the proofs of Lemmas~\ref{lem:cer:reduce}--\ref{lem:cer:negative} can be skipped. \paragraph{The conditioned representation.} Suppose a sequential projective representation reproduces, from $e/3$, the probability of every word of length at most seven that begins with ``yes'' on $A$. Pass to its real cyclic representation (Lemma~\ref{lem:bri:gns}): a real Hilbert space $\mathcal K$, orthogonal projections on it, written again $P_A,\dots,P_D$, and a unit vector $\Omega$ with $\Pr(w)=\|P_w\Omega\|^2$. Since $\Pr(A)=\frac13>0$, the vector $\xi=P_A\Omega/\|P_A\Omega\|$ is a unit vector with $P_A\xi=\xi$, and since $U_{p_A}(e/3)=\frac13p_A$, \begin{equation}\label{eq:cer:H} \|P_u\xi\|^2=3\Pr\nolimits_{e/3}(A,u)=\Pr\nolimits_{p_A}(u)\qquad\text{for every word $u$ of length at most six.} \end{equation} A representation that reproduces from the pure state $p_A$ every word of length at most six gives \eqref{eq:cer:H} directly, with $\xi=\Omega$: there $\|(1-P_A)\Omega\|^2=1-\Pr(A)=0$. Below, \emph{hypothesis (H)} means: a real Hilbert space with orthogonal projections $P_A,\dots,P_D$ and a unit vector $\xi=P_A\xi$ satisfying \eqref{eq:cer:H}. Only this direction of the conditioning is used. All inner products are real. A \emph{walk} is a word $w=w_1\cdots w_n$ in the letters $A,B,C,D$, standing for the operator $P_{w_1}\cdots P_{w_n}$, and \[ L(w)=\ip{\xi}{P_{w_1}\cdots P_{w_n}\xi}. \] $L$ is unchanged by reversing $w$ (the $P_i$ are self-adjoint) and by deleting a letter $A$ at either end ($P_A\xi=\xi$), and $L$ of the empty walk is $1$. If $X=P_{u_k}\cdots P_{u_1}\xi$ and $Y=P_{v_l}\cdots P_{v_1}\xi$, so that $u=(u_1,\dots,u_k)$ and $v=(v_1,\dots,v_l)$ are their words in chronological order, then \begin{equation}\label{eq:cer:walkrule} \ip XY=L(u_1\cdots u_k\,v_l\cdots v_1): \end{equation} the word of $X$ followed by the reversed word of $Y$, equal neighbours merged ($P_i^2=P_i$). A palindromic walk is a word probability: $L(u_1\cdots u_k\cdots u_1)=\|P_u\xi\|^2=\Pr_{p_A}(u)$. \provisionalref{Short histories}{lem:cer:short}{Same statement}{\leandecl{LengthSeven.WalkForm.shortRules_of_table}; from $e/3$: \leandecl{LengthSeven.lengthSeven_e3_rules}.} \begin{lemma}[Relations at short histories]\label{lem:cer:short} Assume (H). Let $q$ be one of $A,B,C,D$, let $t$ be a word of ``yes'' outcomes (in this section a \emph{history} is such a word), and $\eta=P_qP_t\xi$. Then, with the numbers of Definition~\ref{def:traces}, \begin{align*} &\text{(R1$_t$)}\quad P_qP_vP_q\,\eta=o_{qv}\,\eta\qquad(v\ne q,\ |t|\le3),\\ &\text{(R2$_t$)}\quad P_q(P_vP_w+P_wP_v)P_q\,\eta=2R_{qvw}\,\eta\qquad(v,w,q\text{ distinct},\ |t|\le1),\\ &\text{(R3$_0$)}\quad P_A(P_xP_yP_z+P_zP_yP_x)P_A\,\xi=2\omega_y\,\xi\qquad((x,y,z)\text{ an ordering of }B,C,D), \end{align*} where $\omega_B=\omega_1$, $\omega_C=\omega_2$, $\omega_D=\omega_3$ (the index is the middle letter). In the probabilities from $p_A$, the words used are $(t,q,v,q)$, $(t,q,v)$, $(t,q)$ for (R1$_t$); $(t,q,w,q)$, $(t,q,\bar v,w,\bar v,q)$, $(t,q,v,w,v,q)$, $(t,q,w)$, $(t,q,\bar v,w)$, $(t,q,v,w)$, $(t,q)$ for (R2$_t$); and $(v,\check v,A)$, $(v)$ and the empty word for (R3$_0$), with $v$ one of seventeen words of length at most three. All have length at most six; from $e/3$ they are the same words after a leading ``yes'' on $A$, of length at most seven. \end{lemma} \begin{proof} For a word $v$, the purity operator $z_q(v)=P_qE_vP_q-c\,P_q$ of Lemma~\ref{lem:purity} satisfies $z_q(v)^2=P_qE_vP_qE_vP_q-2cP_qE_vP_q+c^2P_q$. Paired with $P_t\xi$ this gives, exactly as in that proof with $\xi$ in place of $\Omega$, $\|z_q(v)P_t\xi\|^2=\|P_{(t,q,v,\check v,q)}\xi\|^2-2c\|P_{(t,q,v)}\xi\|^2+c^2\|P_{(t,q)}\xi\|^2$. When the three words have length at most six, \eqref{eq:cer:H} replaces these by the probabilities from $p_A$, whose combination vanishes, so $z_q(v)\eta=z_q(v)P_t\xi=0$. (R1$_t$) is this with $v=(v)$ and $c=o_{qv}$; the longest word has length $|t|+3$. (R2$_t$): in any associative algebra $P_vP_w+P_wP_v=P_w-E_{(\bar v,w)}+E_{(v,w)}$, where $E_{(\bar v,w)}=(1-P_v)P_w(1-P_v)$ and $E_{(v,w)}=P_vP_wP_v$. Sandwiching by $P_q$ and applying purity to the three terms replaces them by the constants $o_{qw}$, $\ip{p_q}{U_{e-p_v}p_w}=o_{qw}-2R_{qvw}+o_{qv}o_{vw}$ (from $U_{e-p}=\id-2L_p+U_p$) and $\ip{p_q}{U_{p_v}p_w}=o_{qv}o_{vw}$, whose signed sum is $2R_{qvw}$. The longest words are $(t,q,\bar v,w,\bar v,q)$ and $(t,q,v,w,v,q)$, of length $|t|+5$. The letters $v$ and $w$ may be exchanged. (R3$_0$): by the proof of Lemma~\ref{lem:fp:access}, $P_xP_yP_z+P_zP_yP_x$ is a signed sum, with coefficients $\pm\frac12$, of the seventeen sequential products $E_v$ with $v\in\{y;\ zy,\bar zy;\ xy,\bar xy;\ xyz,\bar xyz,x\bar yz,\bar x\bar yz;\ zxy,\bar zxy,z\bar xy,\bar z\bar xy; \ yxz,\bar yxz,y\bar xz,\bar y\bar xz\}$, and the corresponding constants sum to $2Q_{A;xyz}=2\omega_y$ by \eqref{eq:fp:Q}. Here $q=A$, $t$ is empty and $P_A\xi=\xi$; from $p_A$ a leading $A$ costs nothing, so the words $(A,v,\check v,A)$, $(A,v)$, $(A)$ are $(v,\check v,A)$, $(v)$ and the empty word, of length at most six. \end{proof} Since $L$ is unchanged by reversal, each relation may also be applied to the reversed walk; the history is then read from the other end. \paragraph{The vectors.} For an ordering $s=(x,y,z)$ of $B,C,D$ and distinct $u,v\in\{B,C,D\}$ put \[ a_x=P_x\xi,\quad \eta^H_s=P_zP_yP_x\xi,\quad \eta^C_s=P_A\eta^H_s,\quad \eta^R_s=P_x\eta^H_s,\quad \eta^J_{uv}=P_vP_u\xi . \] Lemma~\ref{lem:cer:gram7} in Appendix~\ref{app:cert:seven} expresses every inner product among these vectors as a polynomial in the thirteen traces. Each entry takes one or two applications of Lemma~\ref{lem:cer:short}; the formulas hold in every representation satisfying (H), for every quadruple, and the octonions enter only through the values of the traces. \paragraph{Twenty-eight path vectors.} For a word $w=(w_1,\dots,w_k)$ in the letters $A,B,C,D$ put $[w]=P_{w_k}\cdots P_{w_1}\xi$, and $[\varnothing]=\xi$. The uniform argument uses the set $\mathcal W_{28}$ of the twenty-eight words (a list, not $\mathcal W_L$ for $L=28$), written without commas, \begin{equation}\label{eq:cer:W28} \begin{array}{ll} \text{ending in }A: & \varnothing,\ CBA,\ DCA,\ BDA,\ DCBA,\ CDBA,\ DBCA;\\ \text{ending in }B: & B,\ CB,\ DB,\ CDAB,\ DCB,\ CDB,\ BDCB;\\ \text{ending in }C: & C,\ DC,\ BC,\ DBAC,\ BDC,\ DBC,\ CBDC;\\ \text{ending in }D: & D,\ BD,\ CD,\ BCAD,\ CBD,\ BCD,\ DCBD. \end{array} \end{equation} The words ending in $C$ and $D$ arise from those ending in $B$ by the cyclic relabelling $B\to C\to D\to B$. The $[w]$ are $\xi$, the $a_x$, all six $\eta^J_{uv}$ and $\eta^H_s$, three each of the $\eta^C_s$ and $\eta^R_s$, and the six vectors $[x,z,A]$, $[x,y,A,z]$ with $(x,y,z)$ a cyclic order of $B,C,D$. \paragraph{Walk spaces.} The relations of Lemma~\ref{lem:cer:short} are used below in a second, indefinite space, so the reduction of walks is stated without positivity. A \emph{walk space} is a real vector space $\mathcal Y$ with a symmetric bilinear form $b$, linear maps $P_A,\dots,P_D$ with $P_i^2=P_i$ and $b(P_iu,v)=b(u,P_iv)$, and a vector $\xi=P_A\xi$ with $b(\xi,\xi)=1$. Walks, histories, $[t]=P_{t_r}\cdots P_{t_1}\xi$ and $L(w)=b(\xi,P_{w_1}\cdots P_{w_n}\xi)$ are defined as for (H). As there, $L$ is unchanged by reversal, by merging equal neighbours and by deleting a letter $A$ at either end, and $b([u],[v])=L(u\,v^{\mathrm{rev}})$. A walk is \emph{reduced} (for walks) if it has no equal neighbours and neither begins nor ends with $A$. For a walk $\alpha\,qvq\,\beta$ with $v\ne q$, and for $\alpha\,qvwq\,\beta$ with $q,v,w$ distinct, the relations (R1$_t$) and (R2$_t$) with $t$ the word of $\beta$ read from right to left give \begin{equation}\label{eq:cer:steps} L(\alpha\,qvq\,\beta)=o_{qv}\,L(\alpha\,q\,\beta),\qquad L(\alpha\,qvwq\,\beta)=2R_{qvw}\,L(\alpha\,q\,\beta)-L(\alpha\,qwvq\,\beta). \end{equation} The same identities applied to the reversed walk use as history the part $\alpha$, read from left to right; the step is then said to act \emph{from the left}. (R3$_0$) gives, for $\alpha$ empty or ending in $A$, \begin{equation}\label{eq:cer:R3step} L(\alpha\,xyz)=2\omega_y\,L(\alpha)-L(\alpha\,zyx). \end{equation} \leannote{hand}{} \begin{lemma}[Reduction of walks]\label{lem:cer:reduce} Fix thirteen real numbers $o_{ij}=o_{ji}$, $R_{ijk}$ ($i,j,k$ distinct letters, $R_{ijk}$ symmetric in its indices) and $\omega_B,\omega_C,\omega_D$. Let a walk space satisfy, for all distinct $x,y,z\in\{B,C,D\}$, \begin{enumerate}[label=(\alph*),leftmargin=2em] \item $L(x)=o_{Ax}$ and $L(xyAzyx)=o_{Ax}o_{xy}R_{Ayz}$; \item the vector identities (R1$_t$) for $|t|\le2$, (R2$_t$) for $|t|\le1$ and (R3$_0$) of Lemma~\ref{lem:cer:short}, for all histories $t$ in the letters $A,B,C,D$, with these thirteen numbers in place of the traces. \end{enumerate} Then for every walk $w$ of length at most eight there is a polynomial $\Upsilon(w)$ with integer coefficients in the thirteen numbers, depending only on $w$ and not on the walk space, with $L(w)=\Upsilon(w)$. In particular $b([u],[v])=\Upsilon(u\,v^{\mathrm{rev}})$ for all words $u,v$ of length at most four. \end{lemma} \begin{proof} Each identity below expresses $L(w)$ through $1$, the values in (a), the numbers $o_{qv}$, $2R_{qvw}$, $2\omega_y$, and values of walks that are shorter or treated earlier in the list. Induction on the length $n\le8$ therefore defines $\Upsilon(w)$. It suffices to take $w$ reduced. Then $L(\varnothing)=1$ and $L(x)=o_{Ax}$; let $n\ge2$, and let $x,y,z$ be distinct letters of $\{B,C,D\}$ throughout. \emph{Shortening.} If $w=xAw'$, then (R1$_0$) at $A$ gives $L(w)=o_{Ax}L(w')$, and likewise at the right end. If $w=\alpha\,qvq\,\beta$, then $|\alpha|+|\beta|=n-3\le5$, so one side has length at most two and the first identity of \eqref{eq:cer:steps}, from that side, shortens $w$. From now on $w$ has neither form. Every letter $A$ then stands in places $3,\dots,n-2$, and every factor of $w$ without $A$ is periodic of the form $xyzxyz\cdots$, since its letters differ from both neighbours and from the letters two places away. \emph{No letter $A$.} For $n=2,3$, (R2$_0$) at $A$ and (R3$_0$) paired with $\xi$ give $2L(xy)=2R_{Axy}$ and $2L(xyz)=2\omega_y$, using reversal. For $n=4$, (R2$_0$) at $x$ gives $2L(xyzx)=2R_{xyz}o_{Ax}$. For $5\le n\le8$, write $w=\alpha\,qvsq$; by periodicity $\alpha$ ends in $s$. The second identity of \eqref{eq:cer:steps} with $\beta$ empty gives $L(w)=2R_{qvs}L(\alpha q)-L(\alpha\,qsvq)$, and $\alpha\,qsvq$ contains $sqs$. \emph{One letter $A$, $n=5$.} By reversal $w=xyAuv$ with $u\in\{x,z\}$: \begin{align*} L(xyAxy)&=2R_{Axy}L(xy)-L(yxAxy), & L(xyAxz)&=2R_{Axy}L(xz)-L(xAyxz),\\ L(xyAzx)&=2R_{Axz}L(xy)-L(xyAxz), & L(xyAzy)&=L(yzAyx), \end{align*} by the step at $A$ from the left, at $x$ from the left, at $A$ at the right end, and reversal; the last walk has the second form with the letters renamed. \emph{One letter $A$, $n=6$.} By reversal $w=xyAuvs$ with $\{u,v,s\}=\{x,y,z\}$ and $u\ne y$. The walk $xyAzyx$ is given by (a). For the other three, \eqref{eq:cer:R3step} at the right end gives $L(xyAuvs)=2\omega_vL(xy)-L(xyAsvu)$; the last walk is $xyAzyx$ when $(u,v,s)=(x,y,z)$, and contains $yAy$ otherwise. \emph{One letter $A$, $n=7$, in place four.} Write $w=abcAuvs$ with $\{a,b,c\}=\{u,v,s\}=\{B,C,D\}$ and $u\ne c$. If $s=c$, \eqref{eq:cer:R3step} at the right end gives $2\omega_vL(abc)-L(abcAcvu)$, and the last walk contains $cAc$. If $u=a$, the same identity from the left gives $2\omega_bL(avs)-L(cbaAavs)$, which contains $aAa$. The remaining walk is $abcAbca$, and from the left $L(abcAbca)=2\omega_bL(bca)-L(cbaAbca)$, and $cbaAbca$, whose last letter is its third, is the case $s=c$ above. \emph{One letter $A$, $n=7$, in place three.} Now $w=xyAuvsu$ with $u\ne y$. If $u=x$, the step at $x$ from the left gives $2R_{Axy}L(xvsx)-L(xAyxvsx)$. Further \begin{gather*} L(xyAzyxz)=2R_{Ayz}L(xyxz)-L(xyzAyxz),\\ L(xyAzxyz)=2R_{xyz}L(xyAz)-L(xyAzyxz), \end{gather*} by the step at $y$ from the left with history $x$, and the step at $z$ at the right end; the walk $xyzAyxz$ has its letter $A$ in place four. \emph{Two letters $A$.} Then $n=8$ and $w=xyA\,m_1m_2\,Auv$ with $m_1\ne y$ and $m_2\ne u$. If $m_1=x$ or $m_2=v$, the step at the outer $A$ from that end produces $xAx$ or $vAv$. If $m_2=y$, the step at $y$ from the left with history $x$ turns the factor $yAm_1y$ into $ym_1Ay$, which produces $AyA$; if $m_1=u$, the step at $u$ from the right with history $v$ produces $AuA$ in the same way. What remains is $xyAzxAyz$, and the step at $A$ at the right end gives $2R_{Ayz}L(xyAzx)-L(xyAzxAzy)$, a walk with $m_1=u$. \emph{One letter $A$, $n=8$, in place three.} Here $w=xyAuvsuv$, and the step at $v$ at the right end gives $2R_{uvs}L(xyAuv)-L(xyAuvusv)$, which contains $uvu$. \emph{One letter $A$, $n=8$, in place four.} Here $w=abcAuvsu$ with $u\ne c$. If $u=a$, \eqref{eq:cer:R3step} from the left gives $2\omega_bL(avsa)-L(cbaAavsa)$, which contains $aAa$. Otherwise \begin{gather*} L(abcAbacb)=2R_{Abc}L(abacb)-L(abAcbacb),\\ L(abcAbcab)=2R_{abc}L(abcAb)-L(abcAbacb), \end{gather*} by the step at $b$ from the left with history $a$, and the step at $b$ at the right end; $abAcbacb$ has its letter $A$ in place three. These cases exhaust the reduced walks of length at most eight: by reversal a single letter $A$ may be placed in the left half, and two letters $A$ in places $3,\dots,n-2$ at distance at least three occur only in places three and six of a walk of length eight. \end{proof} Under (H) a walk space is given by $\mathcal K$, and (a), (b) hold: (b) is Lemma~\ref{lem:cer:short}, $L(x)=\Pr_{p_A}(x)$, and $L(xyAzyx)$ is the polarisation $2\ip{P_AY}{P_AP_zY}=\|P_AY\|^2+\|P_AP_zY\|^2-\|P_A(1-P_z)Y\|^2$ with $Y=P_yP_x\xi$, three word probabilities from $p_A$ of length at most four, evaluated with the one-``no'' value $\ip{p_c}{U_{e-p_b}p_a}=o_{ac}-2R_{abc}+o_{ab}o_{bc}$. \paragraph{The comparison form.} Fix four atoms with $\Acal>0$. By Theorem~\ref{thm:fourpoint} no two coincide or are orthogonal, so every $o_{ab}$ lies in $(0,1)$. For $x\in\{B,C,D\}$ put \[ m_x=1/o_{Ax}>1,\qquad \mu_x=m_x-1>0,\qquad m=m_Bm_Cm_D,\qquad \mu=\mu_B\mu_C\mu_D, \] and $m_A=1$. Number the edges of the triangle $BCD$ as $1=BC$, $2=CD$, $3=DB$, so that edge $a$ and the vertex $A$ span the triangle of the holonomy $h_a$ of Section~\ref{sec:fp:tetra} with base $j=A$ and order $(k,l,m)=(B,C,D)$; for $x\in\{B,C,D\}$ let $a(x)$ be the edge opposite $x$, so $a(B)=2$, $a(C)=3$, $a(D)=1$. With $\omega_B,\omega_C,\omega_D$ as in Lemma~\ref{lem:cer:short}, put \begin{gather*} \hat c_1=R_{ABC}m_Bm_C,\quad \hat c_2=R_{ACD}m_Cm_D,\quad \hat c_3=R_{ADB}m_Dm_B,\quad \hat c_4=R_{BCD}\,m,\\ \hat M_{11}=o_{BC}m_Bm_C-\hat c_1^2,\quad \hat M_{22}=o_{CD}m_Cm_D-\hat c_2^2,\quad \hat M_{33}=o_{BD}m_Dm_B-\hat c_3^2,\\ \hat M_{12}=\hat c_1\hat c_2-\omega_C\,m,\quad \hat M_{23}=\hat c_2\hat c_3-\omega_D\,m,\quad \hat M_{13}=\hat c_1\hat c_3-\omega_B\,m,\\ \hat S=\hat c_1\hat c_2\hat c_3-\hat c_1\hat M_{23}-\hat c_2\hat M_{13}-\hat c_3\hat M_{12},\qquad \hat b=\hat c_4-\hat S,\qquad \hat D=\det\hat M-\hat b^2 . \end{gather*} These are the numbers $c_a$, $M$, $S$, $c_4$ of Section~\ref{sec:fp:tetra} in another normalisation: with $f_1=(o_{BC}m_Bm_C)^{1/2}$, $f_2=(o_{CD}m_Cm_D)^{1/2}$, $f_3=(o_{BD}m_Dm_B)^{1/2}$ one has $\hat c_a=f_ac_a$ for $a\le3$, $\hat M_{ab}=f_af_bM_{ab}$ (for instance $f_1f_2K_{12}=\omega_C\,m$), $\hat S=f_1f_2f_3S$ and $\hat c_4=f_1f_2f_3c_4$. Hence, by Proposition~\ref{prop:Dacal}, \begin{equation}\label{eq:cer:Dhat} \hat D=(f_1f_2f_3)^2\,\Dtet=o_{BC}o_{CD}o_{BD}\,m^2\,\Dtet=m^3\Acal>0 . \end{equation} Moreover $\hat M$ is positive definite: $M$ is the Gram matrix of the imaginary parts of the octonionic holonomies (end of Section~\ref{sec:fp:tetra}), so $\hat M=\operatorname{diag}(f_a)\,M\operatorname{diag}(f_a)$ is positive semidefinite, and $\det\hat M=\hat D+\hat b^2>0$ makes it positive definite. \emph{The algebra.} Let $\mathcal C$ be the eight-dimensional real Clifford algebra of the proof of Lemma~\ref{lem:A:clifford}, generated by $\gamma_1,\gamma_2,\gamma_3$ with $\gamma_k^2=-1$ and $\gamma_k\gamma_l=-\gamma_l\gamma_k$ ($k\ne l$). Write $\hat M=\Lambda\Lambda^{\mathsf T}$ with $\Lambda$ real and invertible, and put $\epsilon_a=\sum_k\Lambda_{ak}\gamma_k$, so that $\epsilon_a\epsilon_b+\epsilon_b\epsilon_a=-2\hat M_{ab}$. The elements $\hat h_a=\hat c_a+\epsilon_a$ ($a=1,2,3$) play the part of the holonomies. As in the proof of Lemma~\ref{lem:A:clifford}, $\hat\Gamma=\frac12(\epsilon_1\epsilon_2\epsilon_3-\epsilon_3\epsilon_2\epsilon_1) =\det(\Lambda)\gamma_1\gamma_2\gamma_3$ is central, $\hat\Gamma^2=\det\hat M$, and $\hat\Gamma=\epsilon_1\epsilon_2\epsilon_3+\hat M_{23}\epsilon_1-\hat M_{13}\epsilon_2+\hat M_{12}\epsilon_3$ by \eqref{eq:fp:volume}. The eight elements \begin{equation}\label{eq:cer:Cbasis} 1,\quad \epsilon_1,\ \epsilon_2,\ \epsilon_3,\quad \chi_1=\epsilon_2\epsilon_3+\hat M_{23},\ \chi_2=\epsilon_3\epsilon_1+\hat M_{13},\ \chi_3=\epsilon_1\epsilon_2+\hat M_{12},\quad \mathfrak g=\hat\Gamma-\hat b \end{equation} form a basis, since they differ from the ordered words in the $\epsilon_a$ by a unitriangular change. For $\beta\in\mathcal C$ write $[\beta]_{\mathfrak g}$ and $[\beta]_{\chi_a}$ for its coefficients of $\mathfrak g$ and $\chi_a$. Multiplying out with the Clifford relations, $\mathfrak g$ occurs in $\epsilon_a$ times a basis element only in \begin{equation}\label{eq:cer:epschi} \epsilon_a\chi_a=\mathfrak g+\hat b+\hat M_{ac}\epsilon_b-\hat M_{ab}\epsilon_c\qquad((a,b,c)\text{ a cyclic order of }1,2,3), \end{equation} so $[\epsilon_a\beta]_{\mathfrak g}=[\beta]_{\chi_a}$ for every $\beta$. \emph{The form on $\mathcal C$.} Let $u\mapsto\bar u$ be the conjugation of $\mathcal C$, the linear anti-automorphism with $\bar\gamma_k=-\gamma_k$; it exists because the defining relations are unchanged by reversing products and changing the signs of the generators. It changes the sign of the $\epsilon_a$ and $\chi_a$ and fixes $1$ and $\mathfrak g$. Let $\vartheta$ be the linear form with $\vartheta(1)=1$ that vanishes on the other seven elements of \eqref{eq:cer:Cbasis}, and put $\ip uv_{\mathcal C}=\vartheta(\bar uv)$. Since $\vartheta(\bar u)=\vartheta(u)$, this form is symmetric and left multiplication by $w$ has adjoint left multiplication by $\bar w$. So each $\epsilon_a$ is skew, $\hat h_a$ has adjoint $\hat h_a^*=\hat c_a-\epsilon_a$, and $\hat h_a\hat h_a^*=\hat h_a^*\hat h_a=\hat c_a^2+\hat M_{aa}$ is a scalar. Only two values of the form are needed: with $\mathfrak g_a=-\epsilon_a\mathfrak g$, \begin{equation}\label{eq:cer:ghost} \ip\beta{\mathfrak g}_{\mathcal C}=\hat D\,[\beta]_{\mathfrak g},\qquad \ip\beta{\mathfrak g_a}_{\mathcal C}=\hat D\,[\beta]_{\chi_a} \qquad(\beta\in\mathcal C). \end{equation} Indeed, $\mathfrak g$ is central and fixed by conjugation, so for $\beta=\epsilon_a$ or $\chi_a$ the element $\bar\beta \mathfrak g$ changes sign under conjugation and $\vartheta(\bar\beta \mathfrak g)=0$. Also $\vartheta(\mathfrak g)=0$ and, since $\vartheta(\hat\Gamma)=\hat b$, $\vartheta(\mathfrak g^2)=\det\hat M-2\hat b^2+\hat b^2=\hat D$. The second identity follows from the first, since $\ip\beta{\mathfrak g_a}_{\mathcal C}=\ip{\epsilon_a\beta}{\mathfrak g}_{\mathcal C}$. \emph{The form on four copies of $\mathcal C$.} Let $\varsigma(\beta)=\beta-2[\beta]_{\mathfrak g}\,\mathfrak g$, the reflection of $\mathcal C$ in the direction $\mathfrak g$; by \eqref{eq:cer:ghost} it is self-adjoint, and $\varsigma^2=1$. On $\mathcal V=\mathcal C^4$, with elements $F=(F_A;F_B;F_C;F_D)$, define the \emph{comparison form} \[ \mathcal Q(F,F')=\sum_{x,y}\ip{F_x}{T_{xy}F'_y}_{\mathcal C}, \] where $T_{AA}=T_{Ax}=T_{xA}=1$, $T_{xx}=m_x$ for $x\in\{B,C,D\}$, and \[ T_{BC}=\hat h_1\varsigma,\quad T_{CB}=\varsigma\hat h_1^*,\quad T_{CD}=\hat h_2\varsigma,\quad T_{DC}=\varsigma\hat h_2^*,\quad T_{DB}=\hat h_3\varsigma,\quad T_{BD}=\varsigma\hat h_3^*, \] the elements of $\mathcal C$ acting by left multiplication. $\mathcal Q$ is symmetric because $T_{yx}=T_{xy}^*$. For a letter $x$ let $P_x$ be the operator on $\mathcal V$ whose $x$-component is $(P_xF)_x=m_x^{-1}\sum_yT_{xy}F_y$ and whose other components vanish. Since $T_{xx}=m_x$, $P_x^2=P_x$, and $\mathcal Q(F,P_xF')=m_x^{-1}\sum_{z,y}\ip{F_z}{T_{zx}T_{xy}F'_y}_{\mathcal C} =\mathcal Q(P_xF,F')$. For a word $w$ put $\hat u(w)=P_{w_k}\cdots P_{w_1}(1;0;0;0)$, the counterpart of $[w]$. So $(\mathcal V,\mathcal Q,P_A,\dots,P_D)$ has the algebraic shape of a representation, but $\mathcal Q$ is indefinite (Lemma~\ref{lem:cer:negative}): it is a bookkeeping device, not a representation. Its blocks are those of an associative model with holonomies $\hat h_a$ at $A$, except for the reflection $\varsigma$ (the \emph{twist}) in the direction $\mathfrak g$ along the three edges of $BCD$. \provisionalref{Comparison form}{lem:cer:tablemodel}{Lean scope narrower}{\leandecl{LengthSeven.TableModel.tableModel}, \leandecl{LengthSeven.TableModel.tableModel_lemmaB}. Lean proves the last sentence for $u,v\in\mathcal W_{28}$, entry by entry and for all traces with $o_{AB}o_{AC}o_{AD}\ne0$, with Lean's Gram table in place of $\Upsilon$; the walk-space assertion is not stated.} \begin{lemma}[The comparison form satisfies the relations]\label{lem:cer:tablemodel} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$. Then $(\mathcal V,\mathcal Q,P_A,\dots,P_D, \hat u(\varnothing))$ is a walk space satisfying (a) and (b) of Lemma~\ref{lem:cer:reduce} with the thirteen traces of the quadruple. Hence $\mathcal Q(\hat u(u),\hat u(v))=\Upsilon(u\,v^{\mathrm{rev}})$ for all words $u,v$ of length at most four. \end{lemma} \begin{proof} The walk-space axioms were noted after the definition of $\mathcal Q$, and $\mathcal Q(\hat u(\varnothing),\hat u(\varnothing))=\vartheta(1)=1$. For $F$ with only an $x$-component, $P_yF$ has only the $y$-component $m_y^{-1}T_{yx}F_x$. Two facts are used throughout: $\varsigma$ fixes every element with no $\mathfrak g$-coefficient, and an element of $\mathcal C$ that is a product of at most two of the $\hat h_a,\hat h_a^*$ has no $\mathfrak g$-coefficient, by \eqref{eq:cer:epschi}. In particular every vector $P_q\hat u(t)$ with $|t|\le1$ has a $q$-component that is a multiple of $1$ or of $T_{qr}1\in\{\hat h_a,\hat h_a^*\}$ ($r\ne q$ in $\{B,C,D\}$), and so lies in the span of $1$ and the $\epsilon_a$. \emph{(R1).} $P_qP_vP_q=m_q^{-1}m_v^{-1}T_{qv}T_{vq}P_q$. If $A\in\{q,v\}$ this is $o_{Ax}P_q$. On the edge $a$ joining $q$ and $v$, $T_{qv}T_{vq}$ is $\hat h_a\varsigma^2\hat h_a^*$ or $\varsigma\hat h_a^*\hat h_a\varsigma$, both equal to $\hat c_a^2+\hat M_{aa}=o_{qv}m_qm_v$. So (R1) holds on all of $\mathcal V$. \emph{(a).} $L(x)=m_x^{-1}\vartheta(1)=o_{Ax}$. Let $y,z$ span the edge $a$ and let $\beta$ be the $y$-component of $\hat u(xy)$, a multiple $(m_xm_y)^{-1}T_{yx}1$ of $\hat h_c$ or $\hat h_c^*$, $c$ the edge $xy$. Then $L(xyAzyx)=\mathcal Q(\hat u(xy),P_AP_z\hat u(xy)) =m_z^{-1}\ip\beta{T_{yz}\beta}_{\mathcal C}$ (using $T_{zy}=T_{yz}^*$), and since $\varsigma\beta=\beta$ and $\varsigma$ is self-adjoint, $\ip\beta{T_{yz}\beta}_{\mathcal C}$ equals $\ip\beta{\hat h_a\beta}_{\mathcal C}$ or $\ip\beta{\hat h_a^*\beta}_{\mathcal C}$, which is $\hat c_a\ip\beta\beta_{\mathcal C}$ because $\epsilon_a$ is skew. With $\ip\beta\beta_{\mathcal C}=o_{xy}/(m_xm_y)$ and $\hat c_a=R_{Ayz}m_ym_z$ this is $o_{Ax}o_{xy}R_{Ayz}$. \emph{(R3$_0$).} The vector $P_A(P_xP_yP_z+P_zP_yP_x)\hat u(\varnothing)$ has only the $A$-component $m^{-1}(T_{xy}T_{yz}+T_{zy}T_{yx})1$. Every $\varsigma$ here acts on a product of at most two factors and can be deleted, which leaves $m^{-1}(\hat h_a\hat h_b+\hat h_b^*\hat h_a^*)$ or $m^{-1}(\hat h_a^*\hat h_b^*+\hat h_b\hat h_a)$ for the two edges $a,b$ at $y$. By the Clifford relations both equal $2m^{-1}(\hat c_a\hat c_b-\hat M_{ab})=2\omega_y$. \emph{(R2$_t$), $|t|\le1$.} Let $\beta$ be the $q$-component of $\eta=P_q\hat u(t)$. The relation reads $(T_{qv}T_{vw}T_{wq}+T_{qw}T_{wv}T_{vq})\beta=2R_{qvw}m_qm_vm_w\beta$, with $m_A=1$. If $q=A$ then $\beta$ is a real multiple of $1$ ($1$, or $m_x^{-1}=o_{Ax}$ for the history $x$), and the relation is linear in $\beta$; on $1$ the left side is $(T_{vw}+T_{wv})1=\hat h_a+\hat h_a^*=2\hat c_a$. If $v=A$ (or $w=A$) the left side is $(T_{wq}+T_{qw})\beta$; since $\hat h_a^*\beta$ is a product of two factors, $\varsigma$ drops out and it equals $(\hat h_a+\hat h_a^*)\beta=2\hat c_a\beta$. In both cases $\hat c_a$ is the required constant. There remains $\{q,v,w\}=\{B,C,D\}$; by the cyclic relabelling take $q=B$. Write $d_B=\hat h_1\hat h_2\hat h_3$ and cyclically $d_C$, $d_D$, and $n_a=\hat h_a\hat h_a^*$. The Hermitian part of each $d_x$ is $\hat c_4+\mathfrak g$, by the expansion after Theorem~\ref{thm:tetra}, so $[d_x]_{\mathfrak g}=1$ and $\varsigma(\bar d_x)=2\hat c_4-d_x$. The left side is $\hat h_1\varsigma\hat h_2\varsigma\hat h_3\varsigma\beta+\varsigma\hat h_3^*\varsigma\hat h_2^* \varsigma\hat h_1^*\beta$, and $\beta$ is a multiple of $1$, $\hat h_3^*$ or $\hat h_1$ (histories $\varnothing$, $A$ or $B$; $D$; $C$): \begin{align*} \beta=1:&\quad d_B+\varsigma(\bar d_B)=2\hat c_4,\\ \beta=\hat h_3^*:&\quad n_3\hat h_1\hat h_2+\varsigma\bigl(\hat h_3^*\varsigma(\bar d_D)\bigr) =n_3\hat h_1\hat h_2+2\hat c_4\hat h_3^*-n_3\hat h_1\hat h_2=2\hat c_4\hat h_3^*,\\ \beta=\hat h_1:&\quad \hat h_1\varsigma(d_C)+n_1\hat h_3^*\hat h_2^* =2\hat c_4\hat h_1-n_1\hat h_3^*\hat h_2^*+n_1\hat h_3^*\hat h_2^*=2\hat c_4\hat h_1. \end{align*} In the second line the outer $\varsigma$ acts on $2\hat c_4\hat h_3^*-n_3\hat h_1\hat h_2$, which has no $\mathfrak g$-coefficient; in the third, $\varsigma(d_C)=d_C-2\mathfrak g=2\hat c_4-\bar d_C$ and $\hat h_1\bar d_C=n_1\hat h_3^*\hat h_2^*$. Since $2\hat c_4=2R_{BCD}m$, (R2) holds. The last assertion follows from Lemma~\ref{lem:cer:reduce} and $\mathcal Q(\hat u(u),\hat u(v))=L(u\,v^{\mathrm{rev}})$. \end{proof} \provisionalref{Gram transfer}{lem:cer:gram28}{Lean scope narrower}{\leandecl{LengthSeven.LemmaA.lemmaA_of_gram28Hyp}, \leandecl{LengthSeven.LemmaA.lemmaA}, with \leandecl{LengthSeven.TableModel.tableModel}. Lean's conclusion is for $u,v\in\mathcal W_{28}$ only, proved entry by entry rather than through Lemma~\ref{lem:cer:reduce}.} \begin{lemma}[The words fix the Gram matrix]\label{lem:cer:gram28} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$. Let $\mathcal K$ be a real Hilbert space with orthogonal projections $P_A,\dots,P_D$ and a unit vector $\xi=P_A\xi$, and let $\mathcal S$ be a set of words with $\|P_u\xi\|^2=\Pr_{p_A}(u)$ for every $u\in\mathcal S$. Suppose that \begin{enumerate}[label=(\roman*),leftmargin=2em] \item $\mathcal S$ contains every word of length at most five; \item for every history $t$ of length one and distinct $q,v,w$, $\mathcal S$ contains the seven words of the instance (R2$_t$) of Lemma~\ref{lem:cer:short} for at least one choice of the letter that appears twice in its longest word; \item $\eta^C_{xyz}+\eta^C_{zyx}=2\omega_y\,\xi$ for every ordering $(x,y,z)$ of $B,C,D$. \end{enumerate} Then $\ip{[u]}{[v]}=\mathcal Q(\hat u(u),\hat u(v))$ for all words $u,v$ of length at most four, in particular for all $u,v\in\mathcal W_{28}$. Hypothesis (H) implies (i)--(iii), by Lemma~\ref{lem:cer:short}. \end{lemma} \begin{proof} The proof of Lemma~\ref{lem:cer:short} applies to each relation instance whose words lie in $\mathcal S$: by (i) this covers (R1$_t$) with $|t|\le2$ and (R2$_0$), by (ii) the instances of (R2$_t$) with $|t|=1$, and (iii) is (R3$_0$). The words behind (a) of Lemma~\ref{lem:cer:reduce} have length at most four. So $\mathcal K$ satisfies the hypotheses of Lemma~\ref{lem:cer:reduce}, and $\ip{[u]}{[v]}=\Upsilon(u\,v^{\mathrm{rev}})=\mathcal Q(\hat u(u),\hat u(v))$ by Lemma~\ref{lem:cer:tablemodel}. \end{proof} Lemma~\ref{lem:cer:gram7} is not used in the proof of Theorem~\ref{thm:cer:uniform}. Its derivation uses only words of length at most five, the instances (R2$_t$) with $|t|=1$ and (R3$_0$), so its formulas hold under the hypotheses (i)--(iii) of Lemma~\ref{lem:cer:gram28}. \provisionalref{Three vectors}{lem:cer:negative}{Same statement, different route}{\leandecl{LemmaB.gram_Xi}, \leandecl{LemmaB.Xi_mem_span_uhat}, \leandecl{LemmaB.no_psd_realisation_traces}: the Gram entries are checked as a ring identity modulo the trace relation.} \begin{lemma}[Three vectors with negative Gram determinant]\label{lem:cer:negative} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$, and put $\alpha_x=\mu/\mu_x$. There are vectors $\Xi_B,\Xi_C,\Xi_D$ in the real span of the $\hat u(w)$, $w\in\mathcal W_{28}$, with \[ \mathcal Q(\Xi_x,\Xi_x)=4\hat D\alpha_x(\alpha_x+\mu_x),\qquad \mathcal Q(\Xi_x,\Xi_y)=-4\hat D\mu\,m_z\quad(\{x,y,z\}=\{B,C,D\}). \] The Gram matrix of the $\Xi_x/(2\alpha_x)$ has determinant $-4\hat D^3/\prod_x(1-o_{Ax})<0$. \end{lemma} \begin{proof} Write $\hat h_a'=\hat h_a-1$ and $d_a=\hat h_a'\hat h_a'^*=(\hat c_a-1)^2+\hat M_{aa}$, a scalar. For $F\in\mathcal V$ put $(TF)_x=\sum_yT_{xy}F_y$, so that $\mathcal Q(F',F)=\sum_x\ip{F'_x}{(TF)_x}_{\mathcal C}$; call $TF$ the \emph{source} of $F$. \emph{Sources.} For $u=(u_B,u_C,u_D)\in\mathcal C^3$ let $\langle u\rangle=(-u_B-u_C-u_D;u_B;u_C;u_D)$. Then $(T\langle u\rangle)_A=0$ and $(T\langle u\rangle)_x=(\Theta u)_x$, where $\Theta$ is the $3\times3$ matrix with entries $\Theta_{xx}=\mu_x$ and $\Theta_{xy}=T_{xy}-1$. Let $\Theta^0$ be $\Theta$ with $\varsigma$ replaced by $1$: its off-diagonal entries are $\hat h_1'$ ($BC$), $\hat h_1'^*$ ($CB$), $\hat h_2'$, $\hat h_2'^*$, $\hat h_3'$ ($DB$), $\hat h_3'^*$ ($BD$). Since $\varsigma=1-2[\,\cdot\,]_{\mathfrak g}\,\mathfrak g$ and $[\hat h_a^*\beta]_{\mathfrak g}=\hat c_a[\beta]_{\mathfrak g}-[\beta]_{\chi_a}$ by \eqref{eq:cer:epschi}, the twist changes $\Theta^0$ only through $\mathfrak g$-coefficients: \begin{equation}\label{eq:cer:twist} (\Theta u)_B=(\Theta^0u)_B-2[u_C]_{\mathfrak g}\,\hat h_1\mathfrak g-2\bigl(\hat c_3[u_D]_{\mathfrak g}-[u_D]_{\chi_3}\bigr)\mathfrak g, \end{equation} and cyclically for the components $C$ and $D$. \emph{Where the trace relation enters.} The vector $K_B[\beta]$ below is the column $B$ of the adjugate of $\Theta^0$, with its products ordered so that the $C$- and $D$-components of $\Theta^0K_B[\beta]$ vanish identically; its $B$-component is the cubic determinant of $\Theta^0$, which the trace relation reduces to $2\mathfrak g$. For $\beta\in\mathcal C$ let $K_B[\beta]\in\mathcal C^3$ have components \[ u_B=(\mu_C\mu_D-d_2)\beta,\quad u_C=(\hat h_2'\hat h_3'-\mu_D\hat h_1'^*)\beta,\quad u_D=(\hat h_2'^*\hat h_1'^*-\mu_C\hat h_3')\beta, \] and define $K_C$, $K_D$ by the cyclic relabelling. Because $d_a$ is a scalar, $(\Theta^0K_B[\beta])_C=(\mu_C\mu_D-d_2-\mu_C\mu_D+d_2)\hat h_1'^*\beta=0$, and likewise $(\Theta^0K_B[\beta])_D=0$. The $B$-component is \[ \bigl(\mu-\mu_Dd_1-\mu_Bd_2-\mu_Cd_3+\hat h_1'\hat h_2'\hat h_3'+\hat h_3'^*\hat h_2'^*\hat h_1'^*\bigr)\beta =(\tau+2\mathfrak g)\beta, \] because the two triple products add up to $2\hat S'+2\hat\Gamma$ (in the expansion the terms linear in the $\epsilon_a$ cancel, the quadratic ones give $-2(\hat c_a-1)\hat M_{bc}$, the cubic ones $2\hat\Gamma$), where $\hat S'$ is $\hat S$ with every $\hat c_a$ replaced by $\hat c_a-1$, and $\tau=\mu-\mu_Dd_1-\mu_Bd_2-\mu_Cd_3+2(\hat b+\hat S')$. This is the one place where the trace relation enters: $\tau=0$. Indeed $\hat S'-\hat S=\hat c_1+\hat c_2+\hat c_3-1-m(\omega_B+\omega_C+\omega_D)$ and $d_1=o_{BC}m_Bm_C-2\hat c_1+1$ (cyclically), so \begin{align*} \tau={}&m-\textstyle\sum_{x0$ ($\{x,y,z\}=\{B,C,D\}$), so $u=\operatorname{adj}(G)(1,1,1)^{\mathsf T}$ has positive entries, and $Gu=(\det G)(1,1,1)^{\mathsf T}$ gives $\mathcal Q\bigl(\sum_xu_xE_x,\sum_xu_xE_x\bigr)=\hat D\det G\sum_xu_x<0$. The trace relation is used once, to make $K_x[\beta]$ an exact null direction of the untwisted matrix $\Theta^0$ up to the central element $\mathfrak g$; the twist then turns the vectors $K_x[\epsilon_a]$ into sources concentrated on $\mathfrak g$. \begin{proof}[Proof of Theorem~\ref{thm:cer:uniform}] (a) Either hypothesis gives (H) on the real cyclic representation (Lemma~\ref{lem:bri:gns}), as explained above. By Lemma~\ref{lem:cer:negative} there are reals $c_{x,w}$ with $\Xi_x=\sum_{w\in\mathcal W_{28}}c_{x,w}\hat u(w)$. Put $e_x=\sum_wc_{x,w}[w]$, a vector of the real Hilbert space $\mathcal K$. By Lemma~\ref{lem:cer:gram28} and bilinearity, $\ip{e_x}{e_y}=\sum_{w,w'}c_{x,w}c_{y,w'}\,\mathcal Q(\hat u(w),\hat u(w'))=\mathcal Q(\Xi_x,\Xi_y)$. So the Gram matrix of the real vectors $e_x/(2\alpha_x)$ has negative determinant by Lemma~\ref{lem:cer:negative}, which is impossible for a Gram matrix of vectors in a real inner product space. The lower bound $L_{\min}\ge7$ is Proposition~\ref{prop:cer:universalsix}. (b) Let $\mathcal W$ be the finite set of reduced words of length at most seven beginning with ``yes'' on $A$. The tables on $\mathcal W$ of sequential projective representations are restrictions of tables on the set $\mathcal W_L$ of Lemma~\ref{lem:cer:recon} with $L=7$, which form a compact set (Lemma~\ref{lem:cer:recon} and the remark after it). By (a) this set does not contain the octonionic table $f$ from $e/3$, so the continuous function $g\mapsto\max_{w\in\mathcal W}|g(w)-f(w)|$ has a positive minimum $\delta$ on it. \end{proof} Lemmas~\ref{lem:cer:reduce}, \ref{lem:cer:tablemodel}, \ref{lem:cer:gram28} and~\ref{lem:cer:negative} are proved here in full. For Lemma~\ref{lem:cer:gram7}, used only for the explicit vectors of Theorem~\ref{thm:cer:seven}, Appendix~\ref{app:cert:seven} gives the method and representative entries. The argument gives no explicit value of $\delta$ for a general quadruple, although the vector $u$ of the remark after Lemma~\ref{lem:cer:negative} already carries the contradiction in a single squared norm. The Gram entries of Lemma~\ref{lem:cer:gram28} are reached through purity relations, each read from the squared norm of a purity operator applied to a vector, so an error $\varepsilon$ in the table moves them by an amount bounded only by a multiple of $\sqrt\varepsilon$, with a constant that depends on every step of the reduction in Lemma~\ref{lem:cer:reduce}. An inequality linear in the word probabilities with explicit constants for $\mathsf M$ and $\mathsf S$ is in \companionref{section ``Explicit margins for $\mathsf M$ and $\mathsf S$''}. \provisionalref{M and S}{thm:cer:seven}{Same statement, different route}{\leandecl{BlockB.seven_M_e3_algebraic}, \leandecl{BlockB.seven_S_e3_algebraic}, \leandecl{BlockB.seven_M_pA_algebraic}, \leandecl{BlockB.seven_S_pA_algebraic} and the Hilbert and quaternionic versions: the squared norms of the vectors $\Phi$ are evaluated from the Gram entries of Lemma~\ref{lem:cer:gram28}, not from Lemma~\ref{lem:cer:gram7}.} \begin{theorem}[Words of length seven suffice for $\mathsf M$ and $\mathsf S$]\label{thm:cer:seven} Let $p_A,p_B,p_C,p_D$ be the atoms of $\mathsf M$ or of $\mathsf S$. The model class is (I). No sequential projective representation, over $\R$, $\C$ or $\Ha$, of any dimension, with projections of any rank and any state, reproduces from $e/3$ the probability of every word of length at most seven that begins with ``yes'' on $A$; none reproduces from $p_A$ the probability of every word of length at most six. In particular $L_{\min}\le7$. \end{theorem} The theorem is a case of Theorem~\ref{thm:cer:uniform}(a). Its proof adds a single explicit vector $\Phi=4\kappa_C+10\kappa_R-9\kappa_H$ (for $\mathsf M$), built from signed sums $\kappa_X$ of the vectors $\eta^X_s$, whose squared norm the formulas of Lemma~\ref{lem:cer:gram7} make equal to $-\frac4{27}$; for $\mathsf S$ a similar vector has squared norm $-\frac2{729}$. These calculations are in Appendix~\ref{app:cert:seven}. \provisionalref{Six-step models (M, S)}{prop:horizon}{Lean scope narrower}{\leandecl{MatchingModels.prop_cer_universalsix_c12}, \leandecl{MatchingModels.lMin_eq_seven}. Lean gives a model on $\C^{12}$ through length six for every generating quadruple, but not the equal mixture of $\mathsf Q_\pm$.} \begin{proposition}[Words of length six do not suffice for $\mathsf M$ and $\mathsf S$]\label{prop:horizon} Let $p_A,p_B,p_C,p_D$ be the atoms of $\mathsf M$ or of $\mathsf S$, and let $\mathsf Q_\pm$ be the quaternionic configurations obtained by replacing $v_D$ by $(1,\,j,\,\pm(i+j)/\sqrt2)$. The equal mixture of the standard representations on $\C^6$ of $\mathsf Q_+$ and $\mathsf Q_-$, a sequential projective representation on $\C^{12}$ with four projections of rank four and the state $\Tr/12$ (Appendix~\ref{app:cert:models}), reproduces the octonionic probability from $e/3$ of every word of length at most six. Hence, with Theorem~\ref{thm:cer:seven}, $L_{\min}=7$ for $\mathsf M$ and for $\mathsf S$. \end{proposition} \begin{proof} The lower bound can be proved for all words at once. Put $b=\sqrt2$ for $\mathsf M$, $b=1$ for $\mathsf S$, and $n=1+b^2$. Before normalization the columns of the Albert configuration and of its two quaternionic partners are \[ A=(1,0,0),\quad B=(1,b,0),\quad C=(1,i,1),\quad D=(1,j,d), \] where $d=\ell$ for the Albert configuration and $d=\pm(i+j)/\sqrt2$ for the partners. In all three cases the six transition probabilities and four shape invariants are \[ \begin{gathered} (o_{AB},o_{AC},o_{AD},o_{BC},o_{BD},o_{CD}) =(1/n,1/3,1/3,1/3,1/3,1/3),\\ (R_{ABC},R_{ABD},R_{ACD},R_{BCD})=(1/(3n),1/(3n),1/9,1/9). \end{gathered} \] For example, the 11 and 22 entries of $p_C\jp p_D$ are $1/9$, and its 12 entry is $[-2(i+j)+d(i-j)]/18$. Its real part vanishes because the coefficients $d_i,d_j$ of $i,j$ in $d$ are equal. Pairing with $p_B$ gives $R_{BCD}=1/9$. The double-pair traces are \[ t_1=\frac1{9n},\qquad t_2=\frac1{9n}+\frac{bd_j}{18n},\qquad t_3=\frac1{9n}-\frac{bd_i}{18n}. \] Here is a direct check that also fixes the normalization. The 11,12,13 entries of $p_A\jp p_C$ are $1/3,-i/6,1/6$; those of $p_B\jp p_D$ are \[ \frac1{3n},\qquad \frac b{3n}-\frac j6,\qquad \frac{-d-bjd}{6n}. \] The trace pairing gives $t_2$ since $\operatorname{Re}(jd)=-d_j$. Interchanging $C,D$ gives the formula for $t_3$, while $\operatorname{Re}(d(i-j))=0$ gives $t_1$. Thus the Albert double-pair vector is the arithmetic mean of the two quaternionic vectors. The quaternionic partners have the transition probabilities and shape invariants of the Albert configuration, which has $\Acal>0$. By the argument of Appendix~\ref{app:cert:universalsix} (paragraphs ``Traces with one occurrence of an atom'' and ``Expansion and counting''), every word probability of length at most six from $e/3$ is then one and the same affine function of $(t_1,t_2,t_3)$ for the three configurations, so the midpoint identity proves agreement for every word through length six. The standard embedding $\iota_{\Ha}:\HH\to H_6(\C)$ preserves quadratic maps and satisfies $\Tr\circ\iota_{\Ha}=2\tr$. Each quaternionic block is represented by rank-two projections with state $\Tr/6$. Their equal direct sum gives precisely the stated rank-four projections on $\C^{12}$ and state $\Tr/12$. This proves $L_{\min}\ge7$ for both configurations. The upper bound $L_{\min}\le7$ is Theorem~\ref{thm:cer:seven}. \end{proof} The two blocks have the octonionic transition probabilities and shape invariants, and double-pair traces $\frac1{27}(1,\frac32,\frac12)$, $\frac1{27}(1,\frac12,\frac32)$ for $\mathsf M$ and $\frac1{18}(1,1\pm\frac{\sqrt2}4,1\mp\frac{\sqrt2}4)$ for $\mathsf S$, whose averages are the octonionic values. \paragraph{Why the mixture fails at seven.} If a word $u=(q_1,\ldots,q_n)$ starts and ends with ``yes'', its \emph{round trip} is $u^{\rm rt}=(q_1,\ldots,q_n,q_{n-1},\ldots,q_1)$, of length $2n-1$. For atoms in $\J$ or $\HH$, purity at the two ends and self-adjointness of the updates give $\Pr_{e/3}(u^{\rm rt})=3\Pr_{e/3}(u)^2$ (the same calculation as Lemma~\ref{lem:purity}). Thus, if the two blocks have probabilities $r\pm\varepsilon$ on $u$ and the Albert value is $r$, their equal mixture exceeds the Albert value on $u^{\rm rt}$ by $3\varepsilon^2$. Take $u=(A,\bar C,\bar B,D)$. Expanding $U_{e-p}=\id-2L_p+U_p$ twice, \[ 3\Pr\nolimits_{e/3}(u)=\ip{p_D}{U_{e-p_B}U_{e-p_C}p_A} =4\,t_2+(\text{terms in the $o_{ab}$ and $R_{abc}$}), \] since the only double-pair trace arises from $4\ip{p_D}{p_B\jp(p_C\jp p_A)}=4\ip{p_B\jp p_D}{p_A\jp p_C}$. The two blocks differ from the octonionic values only in $t_2$ and $t_3$, by $\pm\frac1{54}$ for $\mathsf M$ and $\pm\frac{\sqrt2}{72}$ for $\mathsf S$ (the values displayed above), so $\varepsilon=\frac43\cdot\frac1{54}=\frac2{81}$ for $\mathsf M$ and $\varepsilon=\frac43\cdot\frac{\sqrt2}{72} =\frac{\sqrt2}{54}$ for $\mathsf S$. The excesses on $(A,\bar C,\bar B,D,\bar B,\bar C,A)$ are therefore $3\varepsilon^2=\frac4{2187}$ and $\frac1{486}$. This explains the failure of this particular mixture; the exclusion of every representation is Theorem~\ref{thm:cer:uniform}, and the vector of Theorem~\ref{thm:cer:seven} sees the same mechanism at the level of Gram matrices. \begin{remark}[The obstruction is a variance]\label{rem:cer:variance} The formulas of Lemma~\ref{lem:cer:gram7} are affine in $\omega_1,\omega_2,\omega_3$ except for the products of two $\omega$'s in three $\mathrm{CC}$ entries, which come from the squares of two walk values $L(xyz)=\omega_y$. Summing over the rows, the part of $\|\kappa_C\|^2$ quadratic in the $\omega$'s is \[ \sum_s\bigl(-2\omega_y^2+2\omega_y\omega_z+2\omega_x\omega_y\bigr) =4s_\omega^2-8(\omega_1^2+\omega_2^2+\omega_3^2),\qquad s_\omega=\omega_1+\omega_2+\omega_3, \] and since $\ip{\kappa_H}{\kappa_C}=\|\kappa_C\|^2$, $\|\Phi\|^2$ contains it with the factor $16-2\cdot36=-56$. The quaternionic partners $\mathsf Q_\pm$ of $\mathsf M$ have the transition probabilities and shape invariants of $\mathsf M$ and $(\omega_1,\omega_2,\omega_3)=\frac1{27}(1,0,2)$ or $\frac1{27}(1,2,0)$, whose mean is the octonionic $\frac1{27}(1,1,1)$; they do have representations, so the same formulas apply to them. Their affine parts average to the value for $\mathsf M$, and exchanging $C$ and $D$ shows that $\mathsf Q_+$ and $\mathsf Q_-$ give the same value. The quadratic part is $\frac{12}{729}$ for $\mathsf M$ and $-\frac4{729}$ for $\mathsf Q_\pm$, so for the vector $\Phi$ of the proof of Theorem~\ref{thm:cer:seven} \[ \|\Phi\|^2(\mathsf Q_\pm)=-\frac4{27}-56\Bigl(-\frac4{729}-\frac{12}{729}\Bigr)=\frac{788}{729}\ge0 . \] The whole difference between the partners and $\mathsf M$ is the term $-8\sum\omega_r^2$, a variance: for fixed $s_\omega$ it is largest when the $\omega_r$ are equal. In the mixture of Proposition~\ref{prop:horizon} each block has its own $\omega$'s, and the average of their squares exceeds the square of their average; this is the Gram-level form of the excess $3\varepsilon^2$. \end{remark} One sequence of seven tests, such as that of the round trip above, never suffices; two exclude the mixtures of quaternionic quadruples. A \emph{schedule} of length seven is a sequence $s=(i_1,\dots,i_7)$ of tests; its words are the $2^7$ assignments of outcomes to $s$. For a pairing $\pi=\{\{i,j\},\{k,l\}\}$ of $\{A,B,C,D\}$ write $t_\pi=\ip{p_i\jp p_j}{p_k\jp p_l}$, one of the double-pair traces of Definition~\ref{def:traces}. \leannote{hand}{} \begin{proposition}[One schedule does not separate, two defeat quaternionic mixtures]\label{prop:cer:schedule} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$; the model class is (I). \begin{enumerate}[label=(\alph*),leftmargin=2em] \item For every schedule $s$ of length seven, a sequential projective representation on $\C^{12}$ reproduces from $e/3$ every word of length at most six and every word of $s$. Consequently a list of words of length at most seven that no sequential projective representation reproduces from $e/3$ contains words of length seven from at least two schedules. \item Let finitely many quadruples of atoms of $\J$, each lying in a unital subalgebra isomorphic to $\HH$ (complex and real quadruples included), be represented by the standard representations on $\C^6$ with the state $\Tr/6$, and mix these blocks with positive weights. The mixture does not reproduce from $e/3$ both every word of length at most five and the two words $(A,\bar C,\bar B,D,\bar B,\bar C,A)$ and $(A,\bar B,\bar C,D,\bar C,\bar B,A)$. \end{enumerate} \end{proposition} \begin{proof} (a) Expand a word of $s$ as in the paragraph ``Expansion and counting'' of Appendix~\ref{app:cert:universalsix}. A term has two factors containing all four atoms only if its split positions strictly between $1$ and $7$ are position four alone, and only if $i_1,\dots,i_4$ are distinct and $i_4,\dots,i_7$ are distinct. The parts of those two factors that depend on the double-pair traces are then multiples of $t_\pi$ and $t_{\pi'}$, with $\pi=\{\{i_1,i_2\},\{i_3,i_4\}\}$ and $\pi'=\{\{i_4,i_5\},\{i_6,i_7\}\}$. Let $\Pi$ be the plane of double-pair traces $t=(t_1,t_2,t_3)$ allowed by the trace relation, $t^{\J}\in\Pi$ the octonionic point (Proposition~\ref{prop:cer:quatpair}). If $i_1,\dots,i_4$ or $i_4,\dots,i_7$ repeat a test, let $\Pi_1\subset\Pi$ be any line through $t^{\J}$; otherwise let $\Pi_1$ be the line through $t^{\J}$ on which $t_\pi$ is constant. On $\Pi_1$ every word probability of $s$ is affine in $t$, as are those of length at most six, so the representation of Proposition~\ref{prop:cer:quatpair}(c) for this line reproduces all of them. The consequence holds because a sublist of a list that some representation reproduces is reproduced by it. (b) Suppose the mixture reproduces these words. In each block $\Pr_{e/3}(u^{\rm rt})=3\Pr_{e/3}(u)^2$ for every word $u$ that begins and ends with ``yes'' (paragraph ``Why the mixture fails at seven''), and so does the octonionic table. The mixture matches $\Pr(u)$ and $\Pr(u^{\rm rt})$, so the variance of $\Pr(u)$ over the blocks vanishes and every block has the octonionic value of $\Pr(u)$. For $u=(i,j)$ and $u=(i,\bar j,k)$, with round trips of length three and five and $3\Pr_{e/3}(i,\bar j,k)=o_{ik}-2R_{ijk}+o_{ij}o_{jk}$, this fixes every $o$ and every $R$ in every block. For $u=(A,\bar C,\bar B,D)$ and $u=(A,\bar B,\bar C,D)$, $3\Pr_{e/3}(u)$ is $4t_2$, respectively $4t_1$, plus terms in the $o$ and $R$, which fixes $t_2$ and $t_1$, and Proposition~\ref{prop:fp:tracerel} fixes $t_3$. So each block has the thirteen traces of $p_A,\dots,p_D$, hence $\Acal>0$, while its atoms lie in a quaternionic subalgebra, against Theorem~\ref{thm:fourpoint}(ii). \end{proof} The minimum separating word length depends on the initial state. The analogous minimum from the pure state $p_A$ is six for every quadruple with $\Acal>0$: by Theorem~\ref{thm:cer:uniform}(a) no representation reproduces from $p_A$ every word of length at most six; conversely, conditioning the model of Proposition~\ref{prop:cer:universalsix} on the outcome ``yes'' of $A$ reproduces every word of length at most five from $p_A$, because $\Pr_{p_A}(u)=3\Pr_{e/3}(A,u)$. \begin{remark}[Faithful and tracial states]\label{rem:cer:tracial} For faithful and for corner-tracial states Theorem~\ref{thm:cer:uniform}(a) has a short direct proof. Let $(\mathcal K,P_x,\Omega)$ be the real cyclic representation (Lemma~\ref{lem:bri:gns}), $\mathcal A_0$ the real $*$-algebra generated by the $P_x$, $M_A=P_A\mathcal A_0P_A$ and $\xi=P_A\Omega$. With the notation of Section~\ref{sec:fp:tetra} at $j=A$, the elements $H_a-c_aP_A$, $h_a^*h_a-P_A$, $h_ah_a^*-P_A$, $\frac12(h_ah_b+h_b^*h_a^*)-K_{ab}P_A$ and $\frac12(h_1h_2h_3+(h_1h_2h_3)^*)-c_4P_A$ of $M_A$ annihilate $\xi$, by Lemma~\ref{lem:purity} applied to words of length at most seven that begin with ``yes'' on $A$. If the left kernel of $\xi$ in $M_A$ is two-sided, as it is when the state is faithful on $\mathcal A_0$ or when $X\mapsto\ip\xi{X\xi}$ is tracial on $M_A$ (for instance for the normalised trace), they vanish on the closure of $M_A\xi$, and the proof of Theorem~\ref{thm:tetra} runs there, giving $\Dtet=0$, against $\Acal>0$. The model of Proposition~\ref{prop:horizon} is tracial, which is why it must fail at length seven. \end{remark} \subsection{Words of length eleven for every configuration}\label{sec:cer:eleven} For every configuration with $\Acal>0$, the argument of Section~\ref{sec:fp:tetra} gives an explicit list of words, the same for every such configuration. A finite list of probabilities cannot give the operator relations (R1)--(R3): by Lemma~\ref{lem:purity}, matching the three words $(t,q,v,\check v,q)$, $(t,q,v)$, $(t,q)$ kills the single vector $\pi(z_q(v))\pi(P_t)\Omega$. Each relation is therefore used only on vectors on which it has been tested. Each of (R1), (R2), (R3) is a signed sum of purity operators $z_q(v)$ with $|v|\le1,2,3$ respectively (proof of Lemma~\ref{lem:fp:access}), so applying it to a vector $\pi(P_t)\Omega$, with $t$ of length $n$ ending in the base of the relation, needs the three words above for each such $z_q(v)$, of lengths up to $n+2$, $n+4$, $n+6$. Fix the base $j=A$ and the order $(k,l,m)=(B,C,D)$. Applying $h_1=P_jP_kP_lP_j/\nu_1$ to a vector whose word ends with $j$ appends the letters $l,k,j$, its adjoint appends $k,l,j$, and similarly for $h_2,h_3$. A vector $\pi(P_t)\Omega$ has \emph{holonomy degree} $d$ if $t$ is $(j)$ followed by $d$ such blocks; there are $1+6+36=43$ words of degree at most two, of length at most seven. Table~\ref{tab:cer:ledger} lists the relation applications that run the proof of Theorem~\ref{thm:tetra} on the single vector $\xi=P_j\Omega$ (notation of Section~\ref{sec:fp:tetra}). The words these applications require, after merging adjacent equal letters and removing duplicates, form a list $\mathcal F_A$: there are $1{,}673$ applications and $2{,}041$ distinct words, and the list is printed in \companionref{appendix ``The eleven-step list''}. \begin{table}[htbp]\centering\footnotesize \caption{Where the eleven letters go (base $j$). A relation (R1), (R2), (R3) applied to a vector whose word has length $n$ and ends in the base of the relation needs words of length up to $n+2$, $n+4$, $n+6$ (one more if the word does not end in the base).} \label{tab:cer:ledger} \begin{tabularx}{\linewidth}{@{}>{\raggedright\arraybackslash}p{0.25\linewidth}>{\raggedright\arraybackslash}p{0.15\linewidth}>{\raggedright\arraybackslash}X r@{}}\toprule Relation needed & Relation used & Vectors & Longest word\\\midrule $H_a=c_a$ & (R2) at $j$ & holonomy degree $\le2$ (length $\le7$) & $11$\\ $h_a^*h_a=h_ah_a^*=1$ & (R1) $\times3$ & holonomy degree $\le2$, and one letter more & $11$\\ Clifford relations \eqref{eq:fp:clifford} & (R1), (R3) at $j$ & holonomy degree $\le1$ (length $\le4$) & $10$\\ fourth face \eqref{eq:fp:face} & (R2) at $k$, (R1) & $P_kP_j\Omega$, $P_mP_kP_j\Omega$, $P_lP_kP_j\Omega$, $\xi$ & $6$\\ norm of $\xi$ & none & $P_j\Omega$ & $1$\\\bottomrule \end{tabularx} \end{table} \provisionalref{Length eleven}{thm:eleven}{Different form}{\leandecl{BlockB.thm_eleven_with_overlaps}, which joins \leandecl{BlockB.thm_eleven_full} and the side assertion \leandecl{MatchingModels.overlap_pos_lt_one_of_fourPoint_pos}; the counts: \leandecl{Ends.Eleven.elevenBase_card}, \leandecl{Ends.Eleven.elevenBase_hist}, \leandecl{BlockB.elevenUnion_card}. In Lean the list is generated from an encoding of the 1,673 relation applications behind Table~\ref{tab:cer:ledger} (\leandecl{BlockB.elevenLedgerWords_eq_elevenBase}).} \begin{theorem}[Words of length eleven suffice]\label{thm:eleven} Let $\mathcal F_A$ be the list of $2{,}041$ reduced words in four tests $A,B,C,D$ printed in \companionref{appendix ``The eleven-step list''}, the words required by the relation applications summarised in Table~\ref{tab:cer:ledger}; each begins with ``yes'' on $A$ and has length at most eleven. Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$, equivalently (Theorem~\ref{thm:fourpoint}) atoms generating $\J$; then all six transition probabilities lie in $(0,1)$. The model class is (I). \begin{enumerate}[label=(\alph*),leftmargin=2em] \item No sequential projective representation reproduces $\Pr_{e/3}(w)$ for all $w\in\mathcal F_A$. \item Let $\mathcal F_B,\mathcal F_C,\mathcal F_D$ be obtained from $\mathcal F_A$ by exchanging the label $A$ with $B$, $C$, $D$, and $\mathcal F=\mathcal F_A\cup\dots\cup\mathcal F_D$ ($8{,}164$ words). For every state $\rho$ of $\J$, no sequential projective representation reproduces $\Pr_\rho(w)$ for all $w\in\mathcal F$. \item There are a linear functional $\Lambda$ on the probabilities of the words of $\mathcal F$ and the empty word, and $\gamma>0$, such that $\Lambda\ge0$ on every sequential projective representation and $\Lambda\le-\gamma$ on the octonionic table from every state. \end{enumerate} The list does not depend on the atoms. Its words number $1$, $3$, $15$, $36$, $32$, $94$, $240$, $180$, $432$, $576$, $432$ by length $1,\dots,11$. \end{theorem} \begin{proof}[Proof outline] The side assertion is Theorem~\ref{thm:fourpoint}: no two of the atoms coincide or are orthogonal. (a) Suppose a representation reproduces $\Pr_{e/3}(w)$ for all $w\in\mathcal F_A$. By Lemma~\ref{lem:purity} each relation application of Table~\ref{tab:cer:ledger} holds on the vector to which it is applied, since its three words lie in $\mathcal F_A$. These applications run the tetrahedron argument (the holonomy form after Theorem~\ref{thm:tetra}) on the single vector $\xi=P_A\Omega$: the Clifford relations \eqref{eq:fp:clifford} and the face relation \eqref{eq:fp:face} give $\Gamma\xi=(c_4-S)\xi$, and a Gram computation for $\xi$, $N_a\xi$ and $N_1N_2N_3\xi$ gives $\|\Gamma\xi\|^2=\det M\,\|\xi\|^2$. Since $\|\xi\|^2=\frac13$, $\Dtet=0$, against $\Acal>0$ and Proposition~\ref{prop:Dacal}. (b) $p_A+p_B+p_C+p_D$ is invertible (proof of Corollary~\ref{cor:cer:everystate}), so some $\ip\rho{p_j}>0$, and (a) for the label $j$ applies to the conditioned state as in that proof. (c) follows from (b) as Corollary~\ref{cor:bri:finite} follows from Theorem~\ref{thm:bridgeJ}, by compactness of the tables of representations on the finite list (Lemma~\ref{lem:cer:recon}) and strict separation. The mechanical parts, the list itself, its count and its length histogram, and the verification that each relation application needs only words of the list, are checked in Lean (Appendix~\ref{app:lean-plan}); a proof written for reading is in \companionref{section ``Words of length eleven for every configuration'' and appendices ``A constructive eleven-step separator'' and ``The eleven-step list''}. \end{proof} \subsection{A positive tetrahedron statistic is not a witness}\label{sec:cer:Dfails} \provisionalref{Dtet not a witness}{prop:Dfails}{Same statement}{\leandecl{MatchingModels.dfails_length_le_two}, \leandecl{MatchingModels.dfails_thirteen_traces}, \leandecl{MatchingModels.dfails_tetra_dtet}, \leandecl{MatchingModels.dfails_purity_defect}, \leandecl{MatchingModels.dfails_length_three_deviations}.} \begin{proposition}[$\Dtet>0$ is not a witness]\label{prop:Dfails} The model class is (I). There is a sequential projective representation on $\C^{12}$, with four projections of rank four and the state $\Tr/12$ (Appendix~\ref{app:cert:models}), whose word probabilities agree with those of $\mathsf S$ from $e/3$ on every word of length at most two, and whose thirteen traces, read from words of length at most four by the formulas of Section~\ref{sec:fp:traces}, are those of $\mathsf S$; so its tetrahedron statistic is $\Dtet=\frac1{12}>0$. The purity operator $T=P_CP_DP_C-\tfrac13P_C$ of the pair $(C,D)$ has \[ \varphi(T^2)=\tfrac4{243}, \] and the model's deviations from $\mathsf S$ at length three occur exactly on the words whose tests are $C,D,C$ or $D,C,D$, each by $\pm\frac4{243}$. \end{proposition} \begin{proof} Use the matrices $C_1,C_2,C_3$ printed in Appendix~\ref{app:cert:models} and put $\zeta=C_1C_2C_3=1_2\otimes\sigma_x$. They are skew-Hermitian, square to $-1$ and anticommute. Thus $\zeta$ is a central self-adjoint involution with two-dimensional $+1$ and $-1$ eigenspaces. The normalized trace $\tau=\Tr_4/4$ vanishes on $\zeta$, on each $C_i$ and on each $C_iC_j$ with $i\ne j$. For the latter two assertions, conjugation by a suitable $C_k$ changes the sign; for $\zeta$ use its displayed tensor form. The four column isometries defining the projections are \[ \begin{gathered} V_A=(1,0,0)^{\mathsf T},\quad V_B=(1,1,0)^{\mathsf T}/\sqrt2,\\ V_C=(1,-C_1,1)^{\mathsf T}/\sqrt3,\quad V_D=(1,-C_2,-C_3)^{\mathsf T}/\sqrt3. \end{gathered} \] where each entry is a $4\times4$ block. They satisfy $V_i^*V_i=1_4$, so $P_i=V_iV_i^*$ are rank-four projections. With the $4\times4$ blocks $K_{ij}=V_i^*V_j$, \[ \begin{gathered} K_{AB}=1/\sqrt2,\quad K_{AC}=K_{AD}=1/\sqrt3,\quad K_{BC}=(1-C_1)/\sqrt6,\quad K_{BD}=(1-C_2)/\sqrt6,\\ K_{CD}=(1-C_1C_2-C_3)/3=(1+(\zeta-1)C_3)/3. \end{gathered} \] For all pairs except $CD$, $K_{ij}K_{ji}=o_{ij}1_4$, with $o_{AB}=1/2$ and $o_{ij}=1/3$ otherwise. The remaining product is \[ K_{CD}K_{DC}=(3-2\zeta)/9. \] Its scalar values in the two sectors are $1/9$ and $5/9$. Their equal-weight mean is $1/3$ and their variance is $4/81$. Consequently all yes--yes probabilities through length two agree with $\mathsf S$; complementation and the common one-test marginals give every other outcome choice. \emph{The thirteen traces.} Trace cyclicity gives, for any cyclic word, \[ 3\operatorname{Re}\varphi(P_{i_1}\cdots P_{i_n}) =\operatorname{Re}\tau(K_{i_1i_2}\cdots K_{i_ni_1}). \] The signed-update formulas defining the thirteen traces therefore give their usual pair, triangle and symmetrized four-product expressions. The shape invariants involving $A$ are $1/6,1/6,1/9$. For $BCD$, the numerator is \[ (1-C_1)(1-C_1C_2-C_3)(1+C_2) =2-\zeta-2C_1C_2-C_3+C_1C_3+C_2C_3. \] Its normalized real trace is two and the denominator is eighteen, giving $R_{BCD}=1/9$. The three $\omega$'s are the real cyclic products with label orders $ACBD$, $ABCD$ and $ACDB$. Their numerators, each over eighteen, are \[ (1+C_1)(1-C_2),\qquad (1-C_1)(1-C_1C_2-C_3),\qquad (1-C_1C_2-C_3)(1+C_2). \] Each has normalized real trace one. Thus all $\omega_r=1/18$, and $2t_1=\omega_2+\omega_3$ and its cyclic counterparts give all $t_r=1/18$. At base $A$ the resulting data have \[ (c_1,c_2,c_3,c_4)=(1/\sqrt2,1/\sqrt3,1/\sqrt2,1/\sqrt3),\qquad M=\operatorname{diag}(1/2,2/3,1/2). \] Hence $S=c_1c_2c_3=1/(2\sqrt3)$ and $\Dtet=1/6-1/12=1/12$. \emph{Purity and the three-letter deviations.} The purity operator is \[ T=V_C(-2\zeta/9)V_C^*,\qquad T^2=(4/81)P_C, \] so $\varphi(T^2)=4/243$. In each sector all pairs satisfy $P_iP_jP_i=\lambda_{ij}P_i$. A three-yes word $i,j,k$ therefore has probability $\lambda_{ij}\lambda_{jk}/3$, while the Albert probability is $o_{ij}o_{jk}/3$. Only $\lambda_{CD}$ varies with the sector. The average differs precisely when both edges are $CD$, namely for $CDC$ and $DCD$, and the difference is $\operatorname{Var}(\lambda_{CD})/3=4/243$. This also determines every yes--no choice without listing words. With the tracial initial state, write the probability as $\langle q_i,U_{q_j}q_k\rangle/3$, using the Albert trace pairing on one side and $\langle x,y\rangle=3\varphi(x\jp y)$ on the matrix side. Complementing either endpoint is linear. Complementing the middle outcome uses $U_{1-p}=I-2L_p+U_p$; the additional terms depend only on the single-test, pair traces and shape invariants, which have just been shown equal. Thus the discrepancy is the three-yes discrepancy times the signs of the two endpoint outcomes ($+1$ for yes, $-1$ for no); the middle outcome does not change it. All eight choices for each of $CDC,DCD$ therefore differ by $\pm4/243$, and no other pattern differs. Consecutive repeated tests reduce to at most two tests or to zero, so unreduced words cause no exception. \end{proof} The model is a mixture of two sectors on which the four projections act as quaternionic configurations whose fourth atoms are rotated in opposite directions; the mixture restores the double-pair traces but cannot keep every conditional state pure. Hence the separating words of Theorem~\ref{thm:cer:uniform} include words with three intermediate outcomes. \subsection{No linear equality at short length}\label{sec:cer:equality} Satisfying every linear equality only places a table in the affine hull of the model tables (the empty word supplies the constant term); belonging to the convex set of model tables is a stronger property, and it is the one that fails at seven (Theorem~\ref{thm:cer:uniform}). The certificates are inequalities. A linear \emph{equality} $\sum_wc_w\Pr(w)=0$, valid in every sequential projective representation and violated by $\J$, would also separate. This is the form an $s$-identity (a Jordan polynomial identity that holds in every special Jordan algebra but not in all Jordan algebras~\cite{McCrimmonIdentities2004}) such as Glennie's identity~\cite{Glennie1966} would take if it could be read as a relation among word statistics. Call $c$ a \emph{projective equality} if $\sum_wc_w\varphi(E_w)=0$ in every sequential projective representation. Let $FJ_n$ be the free unital Jordan algebra on idempotents $p_1,\dots,p_n$, $\varpi:FJ_n\to\R[\mathbb Z_2^{*n}]^+$ the homomorphism with $p_i\mapsto\frac12(1+r_i)$, and $V_n\subseteq FJ_n$ the span of the elements $U_{q_1}\cdots U_{q_k}1$, $q_t\in\{p_i,1-p_i\}$. For a combination $c$ put $f(c)=\sum_wc_wU_{q_1}\cdots U_{q_k}1\in V_n$, so that $\varpi(f(c))=E(c)$. \needspace{6\baselineskip} \provisionalref{Linear equalities}{prop:cer:noequality}{Lean scope narrower}{(a) \leandecl{AppF.prop_cer_noequality_a}, with projective equalities defined by real finite-dimensional vector states and shown equivalent to vanishing on algebraic models in Lean's lowest universe, through which representations enter; (d) \leandecl{MatchingModels.no_classI_linearEq_le_six}. Lean has no (b), (c).} \begin{proposition}[Linear equalities]\label{prop:cer:noequality} Fix $n$ tests; the model class is (I). \begin{enumerate}[label=(\alph*),leftmargin=2em] \item $c$ is a projective equality if and only if $E(c)=0$ in $\R[\mathbb Z_2^{*n}]$; it suffices to test real finite-dimensional representations with vector states. \item Every $c$ with $f(c)=0$ in $FJ_n$ is a projective equality and holds for idempotents of every Jordan algebra and every state. These include the arrow-of-time equalities $\Pr(u,i)+\Pr(u,\bar\imath)=\Pr(u)$ and the repetition equalities $\Pr(u,a,a,v)=\Pr(u,a,v)$ and $\Pr(u,i,\bar\imath,v)=0$, but are not generated by them: for $i\ne j$, \begin{equation}\label{eq:cer:extraeq} \Pr(\bar\imath,j)-\Pr(i,j)-\Pr(j)=\Pr(\bar\jmath,i)-\Pr(j,i)-\Pr(i),\qquad \Pr(i,j,\bar\imath)=\Pr(i,\bar\jmath,\bar\imath). \end{equation} \item Let $p_1,\dots,p_n$ be idempotents of $\J$. A projective equality violated by their table from some state exists if and only if some $x\in V_n$ has $\varpi(x)=0$ and $x(p_1,\dots,p_n)\ne0$. If the $p_i$ generate $\J$, some $x\in FJ_n$ has these properties; so if $V_n=FJ_n$, such an equality exists. \item For any four atoms of $\J$ with $\Acal>0$, no projective equality supported on words of length at most six is violated from $e/3$. \end{enumerate} \end{proposition} \begin{proof} Parts (a) and (b) are proved in Appendix~\ref{app:cert:noeq}. (c) The octonionic table is $\Pr_\rho(w)=\ip\rho{f(w)(p)}$, so $c$ is violated for some $\rho$ iff $f(c)(p)\ne0$, and by (a) it is a projective equality iff $\varpi(f(c))=0$. If the atoms generate $\J$ and every $x\in\ker\varpi$ had $x(p)=0$, evaluation at $p$, which maps $FJ_n$ onto $\J$, would factor through the special Jordan algebra $\varpi(FJ_n)$; but $\J$ is not a homomorphic image of a special Jordan algebra (Albert and Paige~\cite{AlbertPaige1959}; see also Glennie's proof~\cite[Thm.~4(a), pp.~58--59]{Glennie1966}). (d) For four atoms with $\Acal>0$, the probabilities from $e/3$ of the words of length at most six are reproduced by the representation on $\C^{12}$ of Proposition~\ref{prop:cer:universalsix} (for $\mathsf M$ and $\mathsf S$ also by that of Proposition~\ref{prop:horizon}), and every sequential projective representation satisfies every projective equality. \end{proof} \provisionalref{Five steps}{rem:cer:fivestep}{Same statement}{\leandecl{HorizonFive.horizonFive_projectiveEquality}, assuming only that $c$ vanishes on real finite-dimensional representations with vector states; the form with $E(c)=0$: \leandecl{HorizonFive.horizonFiveClause}.} \begin{remark}[Equalities through five steps]\label{rem:cer:fivestep} For any four atoms of $\J$, in particular those of $\mathsf S$, no projective equality supported on words of length at most five is violated from any state of $\J$. By Proposition~\ref{prop:cer:noequality}(a) this says that every combination $c$ supported on words of length at most five with $E(c)=0$ has $\sum_wc_w\Pr_\rho(w)=0$ for every state $\rho$ of $\J$. It is proved in \companionref{section ``No equality through five steps'' and appendix ``Equalities through five steps''}, and it is machine-checked in both forms (Appendix~\ref{app:lean-plan}). Both proofs show $f(c)(p)=0$, and the idea is as follows. Each word becomes a combination of expressions built from the operators $2L_{r_i}$ and $U_{r_i}$ of the symmetries $r_i=2p_i-e$ (elements with square $e$). Those of degree nine are single reflections with linearly independent special values, so they drop out. Those of degree eight are products of two reflections, and either cancel class by class or contract, without changing their value in $\J$, to lower degree. Below degree eight, a top component involving all four atoms has one atom occurring only once; the other three lie in one complex subalgebra (Lemma~\ref{lem:A:three}), and the associative action of its matrix algebra on the octonionic matrices carries a relation that holds in the group algebra over to $\J$. The needed lower-degree corrections are written down as Jordan polynomials. \end{remark} For general instruments, with independent settings at each time step, the arrow-of-time equalities are the only linear equalities (Clemente and Kofler~\cite{ClementeKofler2016}; Hoffmann, Spee, G\"uhne and Budroni~\cite{HSGB2018}); part (b) shows that a fixed projection per test with the L\"uders rule forces more. Whether a separating equality exists at greater length is open; by (c) it is tied to whether the elements $U_{q_1}\cdots U_{q_k}1$ span the free Jordan algebra on idempotents. An $s$-identity outside $V_n$ has no reading as a combination of word probabilities. Near a complex subalgebra every such equality is weak: its residual vanishes to second order, and states in the subalgebra see it only at third order; this is proved in \companionref{section ``The second-order neighbourhood of a complex subalgebra''}. No link is established here between the length seven of Theorem~\ref{thm:cer:uniform} and the degree eight of Glennie's identity. They are facts of different kinds about the same quadruples. The obstruction of Theorem~\ref{thm:cer:uniform} is one of positivity (a purity operator with three intermediate outcomes needs words of length $1+3+2+1=7$), and every such obstruction needs four tests (Proposition~\ref{prop:three}). Theorem~\ref{thm:glennie} (Appendix~\ref{sec:tet}) is of equality type: a Jordan polynomial of degree two in each of four variables that vanishes in every special Jordan algebra is nonzero at the four symmetries exactly when $\Acal>0$. Whether that polynomial is the value $f(c)$ of a combination $c$ of word probabilities, and of which length, is not decided here. \subsection{Robustness}\label{sec:cer:robust} A numerical optimisation (not a proof) indicates that the margin of Theorem~\ref{thm:cer:uniform}(b), the least maximal word-probability error on its word list that a representation can attain, is about $6.0\times10^{-5}$ for $\mathsf M$ and about $5\times10^{-5}$ for $\mathsf S$. The proof of Theorem~\ref{thm:cer:uniform}(b) gives no value. A separating inequality with known coefficients gives a proved lower bound, by the next proposition; with the inequality of \companionref{section ``Explicit margins for $\mathsf M$ and $\mathsf S$''} the proved margins are above $5.1\times10^{-12}$ for $\mathsf M$ and above $7.7\times10^{-13}$ for $\mathsf S$. \provisionalref{Robust version}{prop:robust}{Lean scope narrower}{\leandecl{FiniteRealWordFunctional.separation_forces_error}. The first assertion; the second, for a target known within $\delta$, is not stated.} \begin{proposition}[Robust version]\label{prop:robust} Let $\Lambda=\sum_w\lambda_w[w]$ (a \emph{witness}) be supported on words of length at most $L$, with $\Lambda(g)\ge0$ for the table $g$ of every sequential projective representation and $\Lambda(f)\le-\gamma<0$ for a target table $f$, and put $\|\lambda\|_1=\sum_w|\lambda_w|$. Every sequential projective representation has $|g(w)-f(w)|\ge\gamma/\|\lambda\|_1$ on some word $w$ of the support; if $|f'(w)-f(w)|\le\delta$ on the support, it misses $f'$ by at least $\gamma/\|\lambda\|_1-\delta$. \end{proposition} \begin{proof} $0\le\Lambda(g)=\Lambda(f)+\Lambda(g-f)\le-\gamma+\|\lambda\|_1\max|g-f|$, and the triangle inequality. \end{proof} The next result concerns instruments with arbitrary positive branches, provided their disturbance is uniformly controlled on both spectral subspaces of the effect rounded to the projection $1_{[1/2,1]}(A)$; it does not infer that control from the observed word probabilities. \provisionalref{Face calibration}{prop:cer:nearideal}{Lean scope narrower}{\leandecl{calibrated_positive_instrument_word_error_le}. The word bound $n\beta$ in an abstract trace-norm setting, with the leakage bound $\tau$ as a hypothesis; the trace-class Hilbert-space instance and \eqref{eq:cer:nearidealwitness} are not formalized.} \begin{proposition}[Stability under uniform face calibration]\label{prop:cer:nearideal} Let $(\mathcal E_1,\mathcal E_0)$ be bounded positive real-linear maps on the self-adjoint trace-class operators of a real, complex or quaternionic Hilbert space, with trace-preserving sum. Write $A=\mathcal E_1^*(1)$, suppose $\|A-A^2\|\le\varepsilon<\frac14$, and set \[ \tau=\tfrac12(1-\sqrt{1-4\varepsilon}),\qquad P=1_{[1/2,1]}(A),\qquad Q=1-P. \] Suppose, for some $\delta\ge0$, that \begin{equation}\label{eq:cer:facecalibration} \|\mathcal E_1(\sigma)-\sigma\|_1\le\delta\quad(\sigma=P\sigma P),\qquad \|\mathcal E_0(\sigma)-\sigma\|_1\le\delta\quad(\sigma=Q\sigma Q) \end{equation} for every density operator on the indicated subspace. Put \[ b=\frac{\delta+\tau+\sqrt{(\delta-\tau)^2+4\tau}}2,\qquad \beta=\min\{2,b\}. \] Then, for every density operator $\rho$, \begin{equation}\label{eq:cer:nearideal} \|\mathcal E_1(\rho)-P\rho P\|_1\le\beta,\qquad \|\mathcal E_0(\rho)-Q\rho Q\|_1\le\beta. \end{equation} For a family of fixed instruments satisfying these hypotheses with the same $\delta,\varepsilon$, the word probabilities from any fixed density operator differ from those of the rounded projective model by at most $n\beta$ on a word of length $n$. Consequently, with the witness and target of Proposition~\ref{prop:robust}, if the instrument table is within $\eta$ of the target on the witness support, then \begin{equation}\label{eq:cer:nearidealwitness} \eta+L\beta\ge\gamma/\|\lambda\|_1. \end{equation} \end{proposition} \begin{proof} The first step controls the off-diagonal part using positivity alone. For a positive map $\mathcal E$ and $\rho\ge0$, put \[ U=\mathcal E(P\rho P),\quad V=\mathcal E(Q\rho Q),\quad H=\mathcal E(P\rho Q+Q\rho P). \] For every real $t$, positivity gives $t^2U+tH+V=\mathcal E((tP+Q)\rho(tP+Q))\ge0$. Let $F_+,F_-$ be the positive and negative spectral support projections of $H$. Taking the real trace against either $F=F_\pm$ and using the discriminant of this nonnegative scalar quadratic gives $|\Tr(FH)|\le2\sqrt{\Tr(FU)\Tr(FV)}$. This also holds when a quadratic coefficient vanishes. Adding the two inequalities, then applying scalar Cauchy--Schwarz and $F_++F_-\le1$, yields \begin{equation}\label{eq:cer:positivecross} \|H\|_1\le2\sqrt{\Tr U\,\Tr V}. \end{equation} The argument uses only self-adjoint trace-class operators and the real trace, so applies over all three scalar fields and in arbitrary dimension. Now let $\rho$ have trace one and put $p=\Tr(P\rho P)$, $q=1-p$. The spectral bound on $A$ gives $QAQ\le\tau Q$. For $\mathcal E=\mathcal E_1$, therefore, $\Tr U\le p$ and $\Tr V=\Tr(AQ\rho Q)\le\tau q$. By \eqref{eq:cer:facecalibration}, homogeneity and \eqref{eq:cer:positivecross}, \[ \|\mathcal E_1(\rho)-P\rho P\|_1 \le\delta p+\tau q+2\sqrt{\tau pq}\le b. \] The last bound is the largest eigenvalue of $\left(\begin{smallmatrix}\delta&\sqrt\tau\\\sqrt\tau&\tau\end{smallmatrix}\right)$, whose quadratic form is evaluated at $(\sqrt p,\sqrt q)$. Interchanging $P,Q$ and $A,1-A$ gives the other branch. Both outputs being positive with trace at most one, their trace-norm distance is also at most two; this gives $\beta$. Every positive trace-nonincreasing map contracts trace norm on self-adjoint trace-class operators: apply it to the positive and negative parts and use the triangle inequality. By homogeneity, \eqref{eq:cer:nearideal} has upper bound $\beta\Tr X$ on any $X\ge0$. Telescoping the two compositions along a word, with the same initial density operator, thus bounds their output trace-norm difference by $n\beta$. Taking traces gives the word-probability bound, and Proposition~\ref{prop:robust} gives \eqref{eq:cer:nearidealwitness}. \end{proof} The calibration in \eqref{eq:cer:facecalibration} concerns every state on the spectral subspaces of the rounded effect, not merely the exact certainty faces of $A$, a finite set of preparations, or states reached in one experiment. The conclusion is a dimension-independent bound for unnormalised branches and word probabilities from density operators; no bound on normalised postselection or on ancillary extensions is asserted. At $\delta=\varepsilon=0$ it recovers the L\"uders rule. For small errors $b\le\delta+\tau+\sqrt\tau$, so the estimate tends to zero. The word-probability bound also combines directly with Theorem~\ref{thm:cer:uniform}(b), which needs no explicit inequality. Let the atoms satisfy $\Acal>0$, let $\delta_0>0$ be the margin of Theorem~\ref{thm:cer:uniform}(b), and suppose that instruments as in the proposition, from a density operator $\rho$, give a table within $\eta$ of the octonionic table from $e/3$ on the words of length at most seven that begin with ``yes'' on $A$. The rounded projections with the state $\rho$ form a sequential projective representation, which differs from the octonionic table by at least $\delta_0$ on one of these words and from the instrument table by at most $7\beta$ on each. Hence \begin{equation}\label{eq:cer:nearidealuniform} \eta+7\beta\ge\delta_0 . \end{equation} When the effects are exact projections, a match on the words of length at most seven forces some test, on a state where its answer is certain, to change the probability of a later word. \leannote{hand}{} \begin{proposition}[Sharp effects disturb a certain answer]\label{prop:cer:sharpdisturb} Let $p_A,p_B,p_C,p_D$ be atoms of $\J$ with $\Acal>0$, and let $\delta_0>0$ be such that every sequential projective representation differs from the octonionic table from $e/3$ by at least $\delta_0$ on some word of length at most seven beginning with ``yes'' on $A$ (Theorem~\ref{thm:cer:uniform}(b)). On a real, complex or quaternionic Hilbert space of any dimension let $\rho_0$ be a density operator and, for each test $i$, let $(\mathcal E_{i,1},\mathcal E_{i,0})$ be bounded positive real-linear maps on the self-adjoint trace-class operators, with trace-preserving sum and with $\mathcal E_{i,1}^*(1)=P_i$ a projection. Write $\Pr_\sigma(q_1,\dots,q_n)=\Tr(\mathcal E_{q_n}\cdots\mathcal E_{q_1}\sigma)$. If $\Pr_{\rho_0}$ equals the octonionic table on all words of length at most seven beginning with ``yes'' on $A$, then there are an outcome $q$ of some test, a density operator $\sigma$ with $\Pr_\sigma(q)=1$, and a nonempty word $u$ of length at most six with \[ |\Pr\nolimits_\sigma(q,u)-\Pr\nolimits_\sigma(u)|\ge \delta_0/2;\qquad\text{in particular}\quad \|\mathcal E_q(\sigma)-\sigma\|_1\ge \delta_0 . \] \end{proposition} \begin{proof} Put $\mathcal L_{i,1}X=P_iXP_i$ and $\mathcal L_{i,0}X=(1-P_i)X(1-P_i)$. For $X\ge0$, $\mathcal E_{i,1}((1-P_i)X(1-P_i))\ge0$ has trace $\Tr(P_i(1-P_i)X(1-P_i))=0$, so it vanishes, and \eqref{eq:cer:positivecross} then kills $\mathcal E_{i,1}(P_iX(1-P_i)+(1-P_i)XP_i)$. Hence $\mathcal E_q=\mathcal E_q\mathcal L_q$ for every outcome $q$ (for ``no'' by the same argument with $1-P_i=\mathcal E_{i,0}^*(1)$), and $\Tr\mathcal E_q=\Tr\mathcal L_q$. The projections $P_i$ with the state $\Tr(\rho_0\,\cdot\,)$ form a sequential projective representation, so some word $w=(q_1,\dots,q_n)$ of the stated kind has $|\Pr_{\rho_0}(w)-\Tr(\mathcal L_{q_n}\cdots\mathcal L_{q_1}\rho_0)|\ge \delta_0$. Let $Y_k=\mathcal L_{q_k}\cdots\mathcal L_{q_1}\rho_0$, $\Phi_k=\mathcal E_{q_n}\cdots\mathcal E_{q_{k+1}}$ and $h_k=\Tr\Phi_kY_k$. Then $h_0=\Pr_{\rho_0}(w)$, $h_n$ is the projective value, and $h_{n-1}=h_n$ because $\Tr\mathcal E_{q_n}=\Tr\mathcal L_{q_n}$. Since $\mathcal E_{q_k}Y_{k-1}=\mathcal E_{q_k}Y_k$, \[ h_{k-1}-h_k=\Tr\Phi_k\bigl(\mathcal E_{q_k}Y_k-Y_k\bigr),\qquad \Tr Y_k\le\Tr(P_A\rho_0)=\Pr\nolimits_{\rho_0}(A)=\tfrac13 . \] The at most six differences with $1\le k\le n-1$ sum to at least $\delta_0$ in absolute value, so one of them is at least $\delta_0/6$ in absolute value. Then $\Tr Y_k>0$. Put $\sigma=Y_k/\Tr Y_k$, $q=q_k$ and $u=(q_{k+1},\dots,q_n)$, a nonempty word since $k\le n-1$. As $\Tr\mathcal E_{q_k}=\Tr\mathcal L_{q_k}$ and $\mathcal L_{q_k}Y_k=Y_k$, $\Pr_\sigma(q)=\Tr\mathcal L_{q_k}\sigma=1$, and \[ \Pr\nolimits_\sigma(q,u)-\Pr\nolimits_\sigma(u)=\frac{\Tr\Phi_k(\mathcal E_{q_k}Y_k-Y_k)}{\Tr Y_k} =\frac{h_{k-1}-h_k}{\Tr Y_k}. \] Dividing a difference of absolute value at least $\delta_0/6$ by $\Tr Y_k\le\frac13$ gives $|\Pr_\sigma(q,u)-\Pr_\sigma(u)|\ge\delta_0/2$, the first bound. The operator $\mathcal E_q(\sigma)-\sigma$ has trace zero and $0\le\Phi_k^*(1)\le1$, so $|\Tr\Phi_k(\mathcal E_q(\sigma)-\sigma)|\le \frac12\|\mathcal E_q(\sigma)-\sigma\|_1$, which gives the second. \end{proof} So, in the notation of Proposition~\ref{prop:cer:nearideal} with $\varepsilon=0$, such instruments have calibration constant $\delta\ge\delta_0$, whereas \eqref{eq:cer:nearidealuniform} alone gives only $\delta\ge\delta_0/7$. Instruments of this kind that reproduce the octonionic table on every word exist (Corollary~\ref{cor:ins:sharpalbert}; Example~\ref{ex:ins:twopress} shows one disturbed certain answer). \begin{remark}[The disturbed state is not prepared by the instruments]\label{rem:cer:sharpdisturb} Suppose the instruments of Proposition~\ref{prop:cer:sharpdisturb} reproduce the octonionic table from $e/3$ on every finite word. Let $K$ be the trace-norm closure of the convex hull of the states $\tau_v=\mathcal E_v(\rho_0)/\Pr_{\rho_0}(v)$, over all words $v$ with $\Pr_{\rho_0}(v)>0$, the empty word included. If $\tau\in K$ and $\Pr_\tau(q)=1$ for an outcome $q$, then $\Pr_\tau(q,u)=\Pr_\tau(u)$ for every word $u$; so the state $\sigma$ of the proposition is not in $K$. Indeed $\Pr_{\tau_v}(u)=\Pr_{e/3}(v,u)/\Pr_{e/3}(v)=\Pr_{\rho_v}(u)$ for the normalized state $\rho_v$ of $\J$ after $v$, so on a finite convex combination $\tau$ of the $\tau_v$ the table is that of the same combination $\rho(\tau)$ of the $\rho_v$. The elements $\theta(w)$ of \eqref{eq:set:prob} span $\J$ (proof of Proposition~\ref{prop:ins:lower}), so finitely many word probabilities determine $\rho(\tau)$ linearly, and each $\Pr_\tau(u)=\Tr\mathcal E_u(\tau)$ is $1$-Lipschitz in trace norm (proof of Proposition~\ref{prop:cer:nearideal}). Hence, if $\tau_n\to\tau$ with every $\tau_n$ in the convex hull, the states $\rho(\tau_n)$ converge to a state $\rho$ of $\J$ with $\Pr_\rho(u)=\Pr_\tau(u)$ for all $u$. If $\Pr_\tau(q)=1$, then $U_q\rho=\rho$ by Lemma~\ref{lem:set:support}, so $\Pr_\rho(q,u)=\Pr_\rho(u)$. \end{remark} \section{Five tests, every four of them quantum}\label{sec:five} For a whole family of tests, four at a time is not enough. Theorem~\ref{thm:pentad} gives five tests of atoms, no two orthogonal, every four of which have a sequential projective representation while the five together have none. Five always suffices (Corollary~\ref{cor:helly}). Words of length at most eleven already exclude the five tests, by the length-seven argument applied to four atoms obtained from them by a reflection; this bound is proved in \companionref{section ``The reflected quadruple and its budget''} and machine-checked (Appendix~\ref{app:lean-plan}). \subsection{Coordinates}\label{sec:fiv:coord} Let $\mathbb K\in\{\R,\C,\Ha,\Oct\}$ and let $H_3(\mathbb K)$ be the Hermitian $3\times3$ matrices over $\mathbb K$ with $x\jp y=\frac12(xy+yx)$; for $\mathbb K=\Oct$ this is $\J$. The Cayley--Dickson units are $i=e_1$, $j=e_2$, $\ell=e_4$ of Appendix~\ref{app:conv}; the products of $i,j,\ell$ give the remaining units, so the only subalgebra of $\Oct$ containing $1,i,j,\ell$ is $\Oct$. For $x\in\mathbb K$ and matrix places $a\ne b$ let $F_{ab}(x)$ be the matrix with $x$ in place $(a,b)$, $\bar x$ in place $(b,a)$ and zeros elsewhere, so $F_{ba}(x)=F_{ab}(\bar x)$, and let $E_{aa}$ be the diagonal matrix units. Matrix places are written $a,b,c$ here and in Appendix~\ref{app:sub}, since $i$ and $j$ denote octonion units. (In the slot notation of Appendix~\ref{app:A:alb}, $F_{23}(x)=F_1(x)$, $F_{31}(x)=F_2(x)$ and $F_{12}(x)=F_3(x)$.) Then $H_3(\mathbb K)$ is the direct sum of $\R E_{11}$, $\R E_{22}$, $\R E_{33}$ and the three spaces $F_{ab}(\mathbb K)$, and for distinct $a,b,c$ \begin{equation}\label{eq:fiv:mult} F_{ab}(x)\jp F_{bc}(y)=\tfrac12F_{ac}(xy). \end{equation} For a column $v\in\mathbb K^3$ whose entries lie in one associative subalgebra of $\mathbb K$ (automatic if they involve at most two imaginary units, by Artin's theorem) put $p(v)=vv^\dagger/\|v\|^2$; it is a primitive idempotent, and its $(a,b)$ entry is $v_a\bar v_b/\|v\|^2$. For two such columns whose entries lie jointly in one associative subalgebra, the computation takes place in a copy of $H_3(\Ha)$, where $\ip xy=\operatorname{Re}\Tr(xy)$, and gives the transition probability \begin{equation}\label{eq:fiv:overlap} \tr\bigl(p(u)\jp p(v)\bigr)=\frac{|u^\dagger v|^2}{\|u\|^2\|v\|^2}. \end{equation} \provisionalref{Coordinates}{lem:five:coord}{Same statement}{\leandecl{lem_five_coord}, \leandecl{lem_five_coord_gen}, with subalgebras of $H_3(\mathbb K)$ taken inside $\J$.} \begin{lemma}[Matrices over a subalgebra]\label{lem:five:coord} Let $B$ be a unital subalgebra of $H_3(\mathbb K)$ that contains $E_{11},E_{22},E_{33}$, $F_{ac}(1)$ and $F_{bc}(1)$, where $\{a,b,c\}=\{1,2,3\}$. Let $\mathbb F=\{x\in\mathbb K:F_{12}(x)\in B\}$. Then $\mathbb F$ is a unital subalgebra of $\mathbb K$ closed under conjugation, and $B=H_3(\mathbb F)$. If $B=\Jord\langle e,X\rangle$, then $\mathbb F$ is the smallest unital subalgebra of $\mathbb K$ closed under conjugation that contains the off-diagonal entries of all elements of $X$. \end{lemma} The proof is in Appendix~\ref{app:sub}. Two atoms $a\ne b$ of $H_3(\mathbb K)$ that lie under the same rank-two idempotent $f$, that is, in the Peirce space $J_0(q)$ of the primitive idempotent $q=e-f$, determine $f$ (proof in Appendix~\ref{app:sub}): \begin{equation}\label{eq:fiv:support} f=\frac{a+b-2\,a\jp b}{1-\tr(a\jp b)} . \end{equation} The atoms $p\le f$ are called the atoms \emph{on the line} $f$ (the projective line of all atoms orthogonal to $q$, in the sense of~\cite{GPR1978}). \subsection{A five-atom configuration}\label{sec:fiv:config} \provisionalref{Five tests}{thm:pentad}{Same statement}{(i)--(iii): \leandecl{pentadF_isAtom}, \leandecl{pentadF_generatesAlbert}, \leandecl{pentadDrop_generated_finrank}; (a) \leandecl{thm_pentad_a}; (b) \leandecl{LengthSeven.Uniform.pentad_b_algebraic}, \leandecl{FiveTest.thm_pentad_b}.} \begin{theorem}[Five tests, every four quantum]\label{thm:pentad} Consider the five atoms of $\J$ \begin{equation}\label{eq:fiv:pentad} \mathsf F=\bigl(p_0,p_1,p_2,p_3,p_4\bigr) =\bigl(E_{11},\ p(1,1,0),\ p(1,i,1),\ p(1,j,0),\ p(1,\ell,0)\bigr). \end{equation} \begin{enumerate}[label=(\roman*),leftmargin=2em] \item No two are orthogonal: $\tr(p_2\jp p_k)=\frac13$ for $k\ne2$, and the other six transition probabilities equal $\frac12$. \item $\Jord\langle e,p_0,\dots,p_4\rangle=\J$. \item Each four-element subfamily $\mathsf F\setminus\{p_k\}$ lies in a unital subalgebra of $\J$ isomorphic to $\HH$. The algebras generated by $\mathsf F\setminus\{p_k\}$, $k=0,\dots,4$, have dimensions $15,15,6,15,15$. \end{enumerate} For the tests of these atoms and model class~(I) (Definition~\ref{def:idealmodel}): \begin{enumerate}[label=(\alph*),leftmargin=2em] \item for every state $\rho$ of $\J$ and every $k$, the four tests $\mathsf F\setminus\{p_k\}$ have a sequential projective representation on $\C^6$ with a density-matrix state and one projection of rank two per test; \item for no state of $\J$ do the five tests have a sequential projective representation, over $\R$, $\C$ or $\Ha$, in any dimension, with projections of any rank and any state of the representation. \end{enumerate} \end{theorem} \begin{proof} \emph{(i)} Each pair of columns involves at most two of $i,j,\ell$, so \eqref{eq:fiv:overlap} applies. For instance $|(1,1,0)^\dagger(1,i,1)|^2=|1+i|^2=2$ gives $\frac2{2\cdot3}=\frac13$, and $|(1,j,0)^\dagger(1,\ell,0)|^2=|1-j\ell|^2=2$ gives $\frac24=\frac12$; the other pairs are alike. \emph{(ii)} Explicitly, \[ \begin{aligned} p_1&=\tfrac12(E_{11}+E_{22})+\tfrac12F_{12}(1), & p_2&=\tfrac13\bigl(e+F_{12}(-i)+F_{13}(1)+F_{23}(i)\bigr),\\ p_3&=\tfrac12(E_{11}+E_{22})+\tfrac12F_{12}(-j), & p_4&=\tfrac12(E_{11}+E_{22})+\tfrac12F_{12}(-\ell). \end{aligned} \] Let $B$ be the generated algebra. It contains $E_{11}=p_0$, hence $U_{e-E_{11}}p_1=\frac12E_{22}$, hence $E_{22}$ and $E_{33}=e-E_{11}-E_{22}$. Peirce components with respect to this frame give $F_{12}(1)$ (from $p_1$) and $F_{13}(1)$ (from $p_2$). Lemma~\ref{lem:five:coord}, with $c=1$, gives $B=H_3(\mathbb F)$, where $\mathbb F$ contains the entries $i$, $j$, $\ell$. So $\mathbb F=\Oct$ and $B=\J$. \emph{(iii)} The proof is in Appendix~\ref{app:sub}. \emph{(a)} Let $Q\cong\HH$ contain the four atoms, $\psi:\HH\to Q$ an isomorphism, and $\iota:\HH\to H_6(\C)$ the symplectic embedding, a unital Jordan monomorphism. An atom of $\J$ in $Q$ is primitive in $Q$, so $\iota\psi^{-1}$ sends it to a rank-two projection. Restrict $\rho$ to $Q$, transport it along $\iota\psi^{-1}$, and extend it to a density matrix as in Section~\ref{sec:bri:converse}; then every word has its octonionic probability, since $\iota\psi^{-1}$ preserves quadratic maps. \emph{(b)} By (ii) and Theorem~\ref{thm:bridgeJ}. \end{proof} By (ii) and (iii), no proper subfamily of $\mathsf F$ generates $\J$. \subsection{Five tests always suffice}\label{sec:fiv:helly} A family $S$ of primitive idempotents is \emph{irredundant} if no proper subfamily generates the same unital subalgebra. Every subfamily of an irredundant family is irredundant (a redundancy in the subfamily is one in $S$). For a Euclidean Jordan algebra $A$ of rank three let $k_{\max}(A)$ be the largest size of an irredundant family of primitive idempotents that generates $A$. \provisionalref{Irredundant reduction}{lem:five:helly}{Same statement}{\leandecl{lem_five_helly}.} \begin{lemma}[Reduction to irredundant families]\label{lem:five:helly} Let $h$ be the least number such that a set of atoms of $\J$ lies in a proper unital subalgebra whenever each of its subsets of at most $h$ elements does. Then $h=k_{\max}(\J)$. \end{lemma} \begin{proof} If $S$ lies in no proper unital subalgebra, it generates $\J$. The generated algebra is spanned by finitely many products of elements of $S$, so a finite subset generates $\J$, and removing members while the remainder still generates gives an irredundant generating subset, of size at most $k_{\max}(\J)$; so $h\le k_{\max}(\J)$. An irredundant generating family of size $k_{\max}(\J)$ has all its proper subsets in proper subalgebras; so $h\ge k_{\max}(\J)$. \end{proof} \provisionalref{Arity}{prop:arity}{Same statement}{\leandecl{prop_arity}.} \begin{proposition}[Largest irredundant generating families]\label{prop:arity} \[ \bigl(k_{\max}(\HR),\,k_{\max}(\HC),\,k_{\max}(\HH),\,k_{\max}(\J)\bigr)=(3,4,4,5). \] \end{proposition} The lower bounds are explicit families. The proof of the upper bounds, in Appendix~\ref{app:sub}, has two ingredients. First, every proper unital subalgebra of $H_3(\mathbb K)$ lies in a \emph{line subalgebra} $\R q\oplus J_0(q)$, for a primitive idempotent $q$, or in a copy of $H_3(\mathbb K')$, where $\mathbb K'$ is the next smaller of $\R,\C,\Ha,\Oct$; this is Lemma~\ref{lem:five:sub}, proved from the classification. Second, one atom off a line together with several atoms on it generates $H_3(\mathbb F)$, where $\mathbb F$ is generated by one off-diagonal entry for each atom on the line. An induction along $\R,\C,\Ha,\Oct$ combines the two. In the language of combinatorial convexity, five is the \emph{Helly number} of the family of proper subalgebras of $\J$, the number $h$ of Lemma~\ref{lem:five:helly}. The same argument with Proposition~\ref{prop:arity} gives the Helly numbers $3$, $4$, $4$ for the proper subalgebras of $\HR$, $\HC$, $\HH$. \provisionalref{Helly numbers}{cor:helly}{Lean scope narrower}{First sentence: \leandecl{cor_helly_geometric}, \leandecl{cor_helly_special}, and \leandecl{cor_helly_sharp} for five. Second sentence: \leandecl{AlbertChallenge.Bridge.cor_helly_classI}, for spaces in Lean's lowest universe, algebraic models admitted, subfamilies indexed by injections; in algebraic-model form, \leandecl{cor_helly_operational_iff_unconditional}.} \begin{corollary}[Helly numbers]\label{cor:helly} A set $S$ of atoms of $\J$, finite or infinite, lies in a proper (equivalently, special) unital subalgebra of $\J$ if and only if each subset of at most five members does, and five cannot be lowered. For model class~(I): a finite family of tests of atoms of $\J$ has a sequential projective representation (from some, equivalently every, state of $\J$) if and only if each subfamily of at most five of its tests does. \end{corollary} \begin{proof} Lemma~\ref{lem:five:helly} and Proposition~\ref{prop:arity}. A proper unital subalgebra has dimension below $27$, so it is special by the classification. The statement about representations is Theorem~\ref{thm:bridgeJ}, which shows that existence depends only on whether the generated algebra is $\J$. \end{proof} \paragraph{Local and global.}\label{sec:fiv:local} All five four-element subfamilies of $\mathsf F$ have $\Acal=0$, as for five atoms in one copy of $\HH$, so whether five atoms generate $\J$ is not a function of the four-point invariants of their subfamilies. In the marginal problem, locally consistent data on a cover extend for all data exactly when the cover is acyclic~\cite{Vorobev1962,AbramskyBrandenburger2011}, a property of the cover; here the number five is a property of the algebra $\J$. For joint measurability of sharp observables in one Hilbert space the analogous number is two (pairwise commutation), while unsharp observables can be pairwise but not triplewise jointly measurable~\cite{HeinosaariReitznerStano2008}. \section{Where the projection postulate enters}\label{sec:ins} The main theorem compares the octonionic table with model class~(I) (Definition~\ref{def:idealmodel}). Letting the hidden quantum system answer each test with more general instruments shows which part of the projection postulate the theorem uses. For one test of an atom $p$, the state after the first outcome is a multiple of $p$ or lies in $J_0(p)$, later tests of $p$ change nothing, and a two-state hidden Markov model suffices. \subsection{A finite realization with non-L\"uders instruments}\label{sec:ins:sharp} \provisionalref{CP realization}{thm:cp69}{Same statement}{\leandecl{NativeAnalysis.exists_nativeGeneral_real_simulation}; the dimension $17n+1$: \leandecl{NativeAnalysis.nativeTaggedBlock_card}.} \begin{theorem}[A finite completely positive realization of ideal-test statistics with non-L\"uders instruments] \label{thm:cp69} Let $p_1,\dots,p_n$ be atoms of $\J$ in any configuration and $\rho$ any state of $\J$. The model class is (II) (Definition~\ref{def:memorymodel}). There are a density matrix $\sigma$ on $\C^{17n+1}$ and, for each test $j$, completely positive maps $\mathcal E_{\mathrm{yes}_j}$, $\mathcal E_{\mathrm{no}_j}$ on $M_{17n+1}(\C)$ with trace-preserving sum, all given by real matrices, such that \[ \Tr\bigl(\mathcal E_{q_m}\cdots\mathcal E_{q_1}\sigma\bigr)=\Pr\nolimits_\rho(q_1,\dots,q_m) \] for every word. For four tests the dimension is $69$. \end{theorem} In the language of Budroni, Fagundes and Kleinmann~\cite{BudroniFagundesKleinmann2019}, this is a quantum finite-state machine of dimension $17n+1$ with general instruments. Because it uses one fixed instrument per test, it reproduces the statistics under every rule for choosing the tests, random or adaptive. Tests of idempotents of rank two are tests of atoms with relabelled outcomes, and ranks $0$ and $3$ are trivial, so every finite family of tests on $\J$ has a finite completely positive realization. \paragraph{The idea.} \begin{enumerate}[label=(\roman*),leftmargin=2em] \item \emph{A ``yes'' resets.} $U_{p_j}x=\ip{p_j}x\,p_j$, so after $\mathrm{yes}_j$ the Albert state is a known multiple of $p_j$; the memory stores one number. \item \emph{A ``no'' lands in a spin factor.} $U_{e-p_j}$ is the orthogonal projection onto the Peirce space $J_0(p_j)\cong H_2(\Oct)$, a ten-dimensional spin factor whose cone is the forward light cone of Minkowski space $\R^{1,9}$. Spin factors are special: they embed in real $16\times16$ matrices through a Clifford representation, a standard fact~\cite{JordanvNWigner1934,FarautKoranyi1994}. \item \emph{Passing between Peirce spaces is a Lorentz similitude unless the atoms are orthogonal.} The restriction of $U_{e-p_j}$ to $J_0(p_i)$ multiplies the Minkowski quadratic form by the transition probability $o_{ij}=\tr(p_i\jp p_j)$. If $o_{ij}>0$ it is carried by a single $16\times16$ spinor matrix; if $o_{ij}=0$ it has rank one and is realized by measuring and preparing. \end{enumerate} The memory has a basis vector $|\mathrm{init}\rangle$ for the initial state, one vector $|P_k\rangle$ per atom $p_k$ (a weight $t$ on it encodes the state $t\,p_k$ after a ``yes''), and one $16$-dimensional block per Peirce space $J_0(p_k)$, holding a spinor-level copy of the state there: $1+n+16n$ dimensions. The proof is in Appendix~\ref{app:mem}. The realization does not embed the full state-and-effect model of $\J$ in quantum theory, preserving all outcome probabilities; such an embedding into finite-dimensional quantum theory is impossible~\cite{GarnerMuller2020,MullerGarner2023}. This obstruction concerns both states and effects, not a map of states alone. It encodes only the states that ideal tests reach after at least one test, and these lie in the special pieces $\R p_j$ and $J_0(p_j)$. \provisionalref{No initial half}{cor:ins:nohalf}{Same statement}{\leandecl{NativeAnalysis.exists_nativeNoHalf_real_simulation}; $68$ for four tests: \leandecl{NativeAnalysis.exists_nativeFourIdempotent_mixed_instrument_68}.} \begin{corollary}[Omitting the initial memory state]\label{cor:ins:nohalf} In Theorem~\ref{thm:cp69}, suppose that for one of the tests $i$ the initial state has the form $\rho=t p_i+y$, with $t\ge0$ and $y\in J_0(p_i)_+$; equivalently, its component in $J_{1/2}(p_i)$ is zero. Then the realization can be made on $\C^{17n}$, with real data. In particular this applies to $\rho=e/3$, giving dimension $68$ for four tests. \end{corollary} \begin{proof} Delete only the initial ray $\C|\mathrm{init}\rangle$ from the construction of Appendix~\ref{app:mem:cp69}. The remaining reset rays $\C|P_k\rangle$ and spinor blocks are invariant under every branch, and the restrictions remain completely positive with trace-preserving sum. Initialize them with $\sigma=t|P_i\rangle\langle P_i|\oplus\mathrm{enc}_i(y)$, with $\mathrm{enc}_i$ the encoding of that appendix and zero on the other blocks. The encoding preserves trace, so $\Tr\sigma=t+\tr y=1$. Each branch advances both summands to the encoding of the corresponding Albert update, by the same block formulas as in that appendix; linearity handles the first test and its induction handles every later word. For $\rho=e/3$, explicitly, $t=1/3$ and $\mathrm{enc}_i((e-p_i)/3)=I_{16}/24$. \end{proof} \begin{remark}[The realization is not projective]\label{rem:ins:notideal} For atoms that generate $\J$, the effect $\mathcal E_{\mathrm{yes}_j}^*(1)$ is not a projection for some $j$: on the block of a Peirce space $J_0(p_i)$, $i\ne j$, with $0= 6}{prop:ins:lower}{Same statement}{\leandecl{native_CP_dimension_at_least_six}; the upper bound $68$: \leandecl{NativeAnalysis.exists_nativeFourIdempotent_mixed_instrument_68}.} \begin{proposition}[Lower bound]\label{prop:ins:lower} Let $p_1,\dots,p_n$ be atoms that generate $\J$, and $\rho=e/3$. Every completely positive realization of their tests (class~(II)) has dimension $d\ge6$. For four such atoms the least dimension $d_{\min}$ satisfies $6\le d_{\min}\le68$. \end{proposition} \begin{proof} In a realization of dimension $d$, the Hankel matrix $H_{u,v}=\Pr(uv)=\Tr\bigl(\mathcal E_v\mathcal E_u\sigma\bigr) =\ip{\mathcal E_v^*(1)}{\mathcal E_u\sigma}_{\mathrm{HS}}$ factors through the real space of Hermitian $d\times d$ matrices, so its rank is at most $d^2$. Bounding the dimension of a realization by the rank of its Hankel matrix is the basic step of the realization theory of recognizable series (Sch\"utzenberger~\cite{Schutzenberger1961}, Fliess~\cite{Fliess1974}). On the Albert side, $U_q$ is self-adjoint, so with $\theta(w)$ as in \eqref{eq:set:prob} \[ H_{u,v}=\tr\bigl(U_{(v)}U_{(u)}\tfrac e3\bigr)=\tfrac13\ip{\theta(v)}{\theta(u^{\mathrm{rev}})}, \] where $U_{(u)}=U_{u_k}\cdots U_{u_1}$ is the composite update along $u=(u_1,\dots,u_k)$ and $u^{\mathrm{rev}}$ is $u$ reversed. So $H$ is the Gram matrix of the $\theta(w)$ with its rows indexed by reversed words, and its rank is $\dim\spn\{\theta(w)\}$. This span contains $e$ and each $p_i$, and it is invariant under $U_{p_i}$ and $U_{e-p_i}$, hence under the reflections $U_{2p_i-e}=2U_{p_i}+2U_{e-p_i}-\id$. It therefore contains every orbit point $q_\alpha$, and by Lemma~\ref{lem:bri:orbit} it is $\J$; that lemma's proof uses only \eqref{eq:bri:lambda}, which holds because atoms of $\J$ are primitive in $B=\J$. So $d^2\ge27$, that is $d\ge6$. The upper bound is Corollary~\ref{cor:ins:nohalf}. \end{proof} \subsection{Sharp effects with arbitrary post-measurement channels}\label{sec:ins:dilation} A realization can also have \emph{sharp} effects, one fixed projection per test, if the update after each outcome is an arbitrary channel. The factorisation of an instrument into a L\"uders instrument followed by outcome-dependent channels, and the Naimark dilation of an effect to a projection, go back to Davies and Lewis~\cite{DaviesLewis1970} and Ozawa~\cite{Ozawa1984}. What is needed here is that the dilation can be made for all tests at once, on one space, with one fixed projection and two fixed channels per test, so that it reproduces \emph{sequences} of tests without a growing record of the past; the short proofs follow. In fixed dimension, instruments of the form ``projective measurement, then a channel'' are a strict subclass of all instruments (Khandelwal and Tavakoli~\cite{KhandelwalTavakoli2025}); in doubled dimension they reproduce the sequential statistics of any two-outcome realization. For a two-outcome instrument $(\mathcal E_+,\mathcal E_-)$ on $M_d(\C)$ write $A=\mathcal E_+^*(1)$, so $0\le A\le1$ and $\mathcal E_-^*(1)=1-A$. \provisionalref{Square-root factor}{lem:ins:factor}{Lean scope narrower}{\leandecl{exists_kraus_channel_factorization}. Existence of the channel, in Kraus form; the formula for $\mathcal F$ is not stated.} \begin{lemma}[Factoring through the square root of the effect]\label{lem:ins:factor} Let $\mathcal E$ be a completely positive map on $M_d(\C)$ with $\mathcal E^*(1)=A$, let $\Pi$ be the support projection of $A$, $A^{-1/2}$ the inverse of $\sqrt A$ on $\operatorname{ran}\Pi$ (extended by $0$), and $\tau_0$ a density matrix. Then \[ \mathcal F(Y)=\mathcal E\bigl(A^{-1/2}YA^{-1/2}\bigr)+\Tr\bigl((1-\Pi)Y\bigr)\,\tau_0 \] is a channel and $\mathcal E(X)=\mathcal F(\sqrt AX\sqrt A)$ for all $X$. \end{lemma} \begin{proof} Write $\mathcal E(X)=\sum_kK_kXK_k^*$ with $\sum_kK_k^*K_k=A$. Each $K_k$ vanishes on $\ker A$, since $\|K_kx\|^2\le\langle x,Ax\rangle$. So $L_k=K_kA^{-1/2}$ satisfies $L_k\sqrt A=K_k\Pi=K_k$ and $\sum_kL_k^*L_k=A^{-1/2}AA^{-1/2}=\Pi$. Hence $\mathcal F$ is completely positive with $\mathcal F^*(1)=\Pi+(1-\Pi)=1$, and $\mathcal F(\sqrt AX\sqrt A)=\sum_kK_kXK_k^*=\mathcal E(X)$ because $(1-\Pi)\sqrt A=0$. \end{proof} \provisionalref{Sharp dilation}{prop:ins:sharpdilation}{Lean scope narrower}{\leandecl{exists_supported_sharpDilation}. The last sentence, on real Kraus operators, is not part of this declaration; for Corollary~\ref{cor:ins:sharpalbert} it is proved.} \begin{proposition}[Sharp effects with arbitrary channels]\label{prop:ins:sharpdilation} Let $\sigma$ and two-outcome instruments $(\mathcal E_{i,+},\mathcal E_{i,-})$, $i=1,\dots,n$, on $M_d(\C)$ form a completely positive realization (Definition~\ref{def:memorymodel}) of some process, and let $A_i=\mathcal E_{i,+}^*(1)$. Put $\mathcal K_2=\C^d\oplus\C^d$ and $\iota x=(x,0)$. For each $i$ there are an orthogonal projection $P_i$ on $\mathcal K_2$ and channels $\Lambda_{i,\pm}$ on $B(\mathcal K_2)$ such that the instruments \[ \widetilde{\mathcal E}_{i,+}(Z)=\Lambda_{i,+}(P_iZP_i),\qquad \widetilde{\mathcal E}_{i,-}(Z)=\Lambda_{i,-}\bigl((1-P_i)Z(1-P_i)\bigr) \] have the following properties. \begin{enumerate}[label=(\alph*),leftmargin=2em] \item $\iota^*P_i\iota=A_i$, the effect of $\widetilde{\mathcal E}_{i,+}$ is $P_i$, and $\widetilde{\mathcal E}_{i,\pm}(\iota X\iota^*)=\iota\mathcal E_{i,\pm}(X)\iota^*$ for all $X\in M_d(\C)$. \item Every output of every $\widetilde{\mathcal E}_{i,\pm}$ lies in $\iota M_d(\C)\iota^*$. \item $\iota\sigma \iota^*$ and the instruments $(\widetilde{\mathcal E}_{i,+},\widetilde{\mathcal E}_{i,-})$ form a completely positive realization of the same process on $\C^{2d}$. \item If each $\mathcal E_{i,+}$ maps positive matrices to matrices supported in the eigenvalue-$1$ space of $A_i$, and each $\mathcal E_{i,-}$ into $\ker A_i$, the channels can be chosen so that the new instruments are repeatable on all states: $\widetilde{\mathcal E}_{i,+}(Z)=P_i\widetilde{\mathcal E}_{i,+}(Z)P_i$ and $\widetilde{\mathcal E}_{i,-}(Z)=(1-P_i)\widetilde{\mathcal E}_{i,-}(Z)(1-P_i)$ for $Z\ge0$. \end{enumerate} If $\sigma$ and all Kraus operators are real matrices, the projections $P_i$ and the channels $\Lambda_{i,\pm}$ can be chosen with real Kraus operators. \end{proposition} \begin{proof} Fix $i$ and drop it from the notation. If $A=0$, put $P=0$, let $\Lambda_+$ be any channel and $\Lambda_-(Z)=\iota\mathcal E_-(\iota^*Z\iota)\iota^*+\Tr((1-\iota \iota^*)Z)\,\iota\tau_0\iota^*$; if $A=1$, exchange the roles of $+$ and $-$. All claims are then immediate, since $\mathcal E_+=0$ when $A=0$. From now on $A\ne0,1$. \emph{The projection.} Since $\sqrt A$ and $\sqrt{1-A}$ commute, \[ U=\begin{pmatrix}\sqrt A&\sqrt{1-A}\\ \sqrt{1-A}&-\sqrt A\end{pmatrix} \] is a self-adjoint unitary on $\mathcal K_2$. Let $\iota'x=(0,x)$, $\iota_+=U\iota$ and $\iota_-=U\iota'$. These are isometries with orthogonal ranges spanning $\mathcal K_2$, so $P=\iota_+\iota_+^*$ is a projection, $1-P=\iota_-\iota_-^*$, and $\iota^*P\iota=A$. Moreover \begin{equation}\label{eq:ins:dilPV} P\iota=U\,\iota \iota^*\,U\iota=\iota_+\sqrt A,\qquad (1-P)\iota=U\,\iota'\iota'^*\,U\iota=\iota_-\sqrt{1-A}. \end{equation} \emph{The channels.} By Lemma~\ref{lem:ins:factor} choose channels $\mathcal F_\pm$ with $\mathcal E_+(X)=\mathcal F_+(\sqrt AX\sqrt A)$ and $\mathcal E_-(X)=\mathcal F_-(\sqrt{1-A}X \sqrt{1-A})$, fix density matrices $\tau_\pm$ on $\mathcal K_2$, and put \[ \begin{aligned} \Lambda_+(Z)&=\iota\mathcal F_+(\iota_+^*Z\iota_+)\iota^*+\Tr\bigl((1-P)Z\bigr)\,\tau_+,\\ \Lambda_-(Z)&=\iota\mathcal F_-(\iota_-^*Z\iota_-)\iota^*+\Tr(PZ)\,\tau_- . \end{aligned} \] Both are completely positive, and $\Tr\Lambda_+(Z)=\Tr(\iota_+\iota_+^*Z)+\Tr((1-P)Z)=\Tr Z$; likewise for $\Lambda_-$. \emph{(a)--(c).} Since $\iota_+^*P=\iota_+^*$ and $(1-P)P=0$, it follows that $\widetilde{\mathcal E}_+(Z)=\iota\mathcal F_+(\iota_+^*Z\iota_+)\iota^*$, and likewise $\widetilde{\mathcal E}_-(Z)=\iota\mathcal F_-(\iota_-^*Z\iota_-)\iota^*$; this is (b). The effects are $P$ and $1-P$. By \eqref{eq:ins:dilPV}, $\iota_+^*\iota X\iota^*\iota_+=\iota_+^*P\iota X\iota^*P\iota_+=\sqrt AX\sqrt A$, so $\widetilde{\mathcal E}_+(\iota X\iota^*)=\iota\mathcal E_+(X)\iota^*$, and the same with $\iota_-$ for the outcome $-$. This is (a). Induction on the word gives $\widetilde{\mathcal E}_w\iota\sigma \iota^*=\iota(\mathcal E_w\sigma)\iota^*$; taking traces gives (c). \emph{(d).} Let $E_1$ and $E_0$ be the eigenvalue-$1$ space and the kernel of $A$; both are nonzero, since $\mathcal E_+\ne0\ne\mathcal E_-$. Take the state $\tau_0$ of $\mathcal F_+$ in $E_1$ and that of $\mathcal F_-$ in $E_0$. Then $\mathcal F_+$ maps positive matrices into matrices supported in $E_1$, and $\mathcal F_-$ into $E_0$. By \eqref{eq:ins:dilPV}, $(1-P)\iota x=\iota_-\sqrt{1-A}\,x=0$ for $x\in E_1$ and $P\iota x=0$ for $x\in E_0$, so $\iota E_1\subseteq \operatorname{ran}P$ and $\iota E_0\subseteq\operatorname{ran}(1-P)$. With (b) this gives (d). The reality statement holds, choosing $\tau_0,\tau_\pm$ real, because $U$, $\iota_+$, $\iota_-$, $P$, $A^{-1/2}$ and the $L_k$ are obtained from the given matrices by real operations and the real spectral theorem. \end{proof} Property~(b) is what makes the construction compose: whichever test comes next finds the state in $\iota M_d(\C)\iota^*$, where its own dilation expects it. \provisionalref{Sharp Albert}{cor:ins:sharpalbert}{Same statement}{\leandecl{NativeAnalysis.exists_nativeSharp_instrument_2_17n_add_one}, \leandecl{NativeAnalysis.exists_nativeNoHalf_sharp_instrument_34n}.} \begin{corollary}[Ideal-test statistics with sharp effects]\label{cor:ins:sharpalbert} Let $p_1,\dots,p_n$ be atoms of $\J$ in any configuration and $\rho$ any state. There are a density matrix on $\C^{2(17n+1)}$ and, for each test $j$, one orthogonal projection $P_j$ and two channels $\Lambda_{j,\pm}$, all real matrices, such that the instruments $Z\mapsto\Lambda_{j,+}(P_jZP_j)$ and $Z\mapsto\Lambda_{j,-}((1-P_j)Z(1-P_j))$ reproduce $\Pr_\rho(w)$ for every word and are repeatable on all states. For four tests the dimension is $138$. Under the initial-state hypothesis of Corollary~\ref{cor:ins:nohalf}, the dimension can instead be $34n$, hence $136$ for four tests, in particular from $e/3$. \end{corollary} \begin{proof} Apply Proposition~\ref{prop:ins:sharpdilation} to the real realization of Theorem~\ref{thm:cp69}; the hypothesis of (d) follows from the block description of Appendix~\ref{app:mem:cp69}. Every instrument there first discards coherences between blocks, so $A_j$ is block diagonal. Every output of $\mathcal E_{\mathrm{yes}_j}$ is a multiple of $|P_j\rangle\langle P_j|$, and $A_j|P_j\rangle=|P_j\rangle$ because the test of $p_j$ is certain to answer ``yes'' on the block of $p_j$. Every output of $\mathcal E_{\mathrm{no}_j}$ lies in the block of $J_0(p_j)$, where $A_j=0$. The same properties hold after deleting the initial ray as in Corollary~\ref{cor:ins:nohalf}, so doubling that realization gives $34n$. \end{proof} In \companionref{section ``Repeatability and a smaller sharp dilation''} the added space has dimension $\max_i\operatorname{rank}(A_i-A_i^2)$ instead of $d$, which lowers these dimensions to $26n-7$ for $n\ge1$, and to $26n-9$ under the initial-state hypothesis ($97$ and $95$ for four tests). \begin{remark}[What fails is ideality]\label{rem:ins:sharpnotideal} Measure the instruments of Corollary~\ref{cor:ins:sharpalbert} against Definition~\ref{def:ideal}. \emph{Sharpness~(Sh) holds.} The faces are the states supported in $\operatorname{ran}P_j$ and in $\operatorname{ran}(1-P_j)$; an effect vanishing on both has $P_jfP_j=(1-P_j)f(1-P_j)=0$, hence $f=0$. \emph{Ideality~(Id) fails.} Let $a=\tr(p_j\jp\rho)$ with $0