SRIC: endogenous semantic witness

Download original file

This is the original TeX manuscript source. Use the paper’s PDF for typeset equations.

Included manuscript sections

• block23_section_body.tex

Read the typeset paper

1 \documentclass[11pt,letterpaper]{article}
2 \usepackage[T1]{fontenc}
3 \usepackage{lmodern,microtype}
4 \usepackage[margin=1in]{geometry}
5 \usepackage{amsmath,amssymb,amsthm,mathtools}
6 \usepackage{booktabs,longtable,array}
7 \usepackage{enumitem,xcolor,hyperref,cleveref,graphicx,float,fancyhdr,setspace,titlesec}
8 \graphicspath{{figures/}{../figures/}}
9 \definecolor{KMBlue}{RGB}{31,70,110}
10 \definecolor{KMGray}{RGB}{85,92,101}
11 \hypersetup{colorlinks=true,linkcolor=KMBlue,citecolor=KMBlue,urlcolor=KMBlue,pdftitle={Endogenous Semantic-Witness Growth and External Latent-Reasoner Pilot},pdfauthor={Komorebi P.},hypertexnames=false}
12 \pagestyle{fancy}\fancyhf{}\fancyhead[L]{\small SRIC Block 23}\fancyhead[R]{\small Komorebi P.}\fancyfoot[C]{\thepage}\setlength{\headheight}{14pt}
13 \titleformat{\section}{\Large\bfseries\color{KMBlue}}{\thesection}{0.75em}{}
14 \titleformat{\subsection}{\large\bfseries\color{KMBlue}}{\thesubsection}{0.65em}{}
15 \setlist[itemize]{leftmargin=1.45em,itemsep=0.25em}\setlist[enumerate]{leftmargin=1.65em,itemsep=0.25em}\setstretch{1.05}
16 \newtheoremstyle{kmplain}{7pt}{7pt}{\itshape}{}{\bfseries\color{KMBlue}}{.}{0.5em}{}
17 \newtheoremstyle{kmdef}{7pt}{7pt}{}{}{\bfseries\color{KMBlue}}{.}{0.5em}{}
18 \newtheoremstyle{kmremark}{6pt}{6pt}{\small}{}{\bfseries\color{KMGray}}{.}{0.5em}{}
19 \theoremstyle{kmplain}\newtheorem{theorem}{Theorem}[section]\newtheorem{proposition}[theorem]{Proposition}\newtheorem{conditionaltheorem}[theorem]{Conditional Theorem}
20 \theoremstyle{kmdef}\newtheorem{definition}[theorem]{Definition}
21 \theoremstyle{kmremark}\newtheorem{warning}[theorem]{Warning}\newtheorem{remark}[theorem]{Remark}
22 \crefname{conditionaltheorem}{Conditional Theorem}{Conditional Theorems}
23 \DeclareMathOperator{\Aut}{Aut}
24 \newcommand{\R}{\mathbb R}\newcommand{\Ucal}{\mathcal U}\newcommand{\Gcal}{\mathcal G}\newcommand{\eps}{\varepsilon}\newcommand{\norm}[1]{\left\lVert #1\right\rVert}
25 \begin{document}
26 \hypersetup{pageanchor=false}
27 \begin{titlepage}
28 \centering\vspace*{1.0in}
29 {\Huge\bfseries\color{KMBlue} Endogenous Semantic-Witness Growth\par}
30 \vspace{0.15in}{\Large and an External Latent-Reasoner Pilot Contract\par}
31 \vspace{0.45in}{\large SRIC Block 23\par}
32 \vfill{\Large Komorebi P.\par}\vspace{0.12in}{\large Version 2.4, September 2026\par}
33 \vfill
34 \begin{minipage}{0.86\textwidth}\small
35 \textbf{Status.} The finite semantic-witness benchmark is complete and reproducible. The third-party Coconut adapter is source locked and protocol tested, but the external checkpoint was not available in the release environment and was not executed. No real-checkpoint interpretability finding is claimed.
36 \end{minipage}
37 \end{titlepage}
38 \hypersetup{pageanchor=true}
39 \pagenumbering{roman}
40 \begin{abstract}
41 We formalize semantic meaning for recurrent latent reasoning as a minimal causal-predictive quotient relative to a declared intervention language. A finite compiler receives trajectories, semantic-preserving twins, anonymous interventions, protected future consequences, and final answers, but no gold intermediate semantic coordinates. Across thirty-six fresh episodes it distinguishes unique identification, identification up to gauge, partial ontology, protected underidentification, and destructive information loss with all certificates correct. An answer-only quotient attains high answer accuracy but fails future and intervention sufficiency. Audit-only counterfactual packet patching supports the selected quotient and rejects answer-only, PCA, and random controls. We also release a source-locked adapter for a public multimode GPT-2 Coconut checkpoint on PrOntoQA. The adapter supports exact trace capture, replay, truncation, permutation, noise replacement, semantic-twin transplantation, and same-answer/different-proof transplantation. The external checkpoint was not present in the release environment, so the real-model track remains an execution-ready protocol rather than an empirical result.
42 \end{abstract}
43 \tableofcontents\clearpage
44 \pagenumbering{arabic}
45 \input{manuscript/block23_section_body.tex}
46 \begin{thebibliography}{9}\footnotesize
47 \bibitem{DILGREN2026}
48 Connor Dilgren and Sarah Wiegreffe,
49 \emph{Are Latent Reasoning Models Easily Interpretable?},
50 arXiv:2604.04902, 2026.
51 \bibitem{SRICB22}
52 Komorebi P.,
53 \emph{SRIC v2.3 Block 22 Correspondence-Free Neuralese Transport from Intervention Orbits},
54 reference implementation and cumulative theory draft, 2026.
55 \end{thebibliography}
56 \end{document}

Back to source controls

Back to SRIC: endogenous semantic witness