| \documentclass[11pt,a4paper,twoside]{article} |
|
|
| |
| \usepackage[utf8]{inputenc} |
| \usepackage[T1]{fontenc} |
| \usepackage{graphicx} |
| \usepackage{booktabs} |
| \usepackage{hyperref} |
| \usepackage{geometry} |
| \usepackage{xcolor} |
| \usepackage{fancyhdr} |
| \usepackage{amsmath} |
| \usepackage{amssymb} |
| \geometry{margin=2.5cm} |
|
|
| |
| \definecolor{gold}{HTML}{D4A017} |
| \definecolor{dark}{HTML}{1a1a2e} |
| \definecolor{accent}{HTML}{16213e} |
|
|
| |
| \hypersetup{ |
| colorlinks=true, |
| linkcolor=accent, |
| urlcolor=accent, |
| citecolor=accent, |
| pdftitle={GraphLang — A Universal Semantic Kernel for Code}, |
| pdfauthor={Josué Argaña Silguero}, |
| pdfsubject={Semantic IR, Code Compression, Cross-Language Analysis}, |
| pdfkeywords={semantic IR, code compression, cross-language, compiler}, |
| } |
|
|
| |
| \pagestyle{fancy} |
| \fancyhf{} |
| \fancyhead[L]{\small GraphLang v1.0.1 — FROZEN} |
| \fancyhead[R]{\small Josué Argaña Silguero} |
| \fancyfoot[C]{\thepage} |
| \renewcommand{\headrulewidth}{0.4pt} |
|
|
| \begin{document} |
|
|
| |
| |
| |
|
|
| \thispagestyle{empty} |
| \begin{center} |
|
|
| \vspace*{3cm} |
|
|
| {\Huge \textbf{GraphLang}} |
|
|
| \vspace{0.5cm} |
|
|
| {\LARGE A Universal Semantic Kernel for Code} |
|
|
| \vspace{0.3cm} |
|
|
| {\Large 29.8x Structural Compression Across 13 Programming Languages} |
|
|
| \vspace{1.5cm} |
|
|
| {\large \textbf{Josué Argaña Silguero}} |
|
|
| \vspace{0.3cm} |
|
|
| {\normalsize Paraguay --- July 28, 2026} |
|
|
| \vspace{0.3cm} |
|
|
| {\small \texttt{josu31.jas@gmail.com}} |
|
|
| \vspace{0.3cm} |
|
|
| {\small \url{https://github.com/cripto-bot/graphlang}} |
|
|
| \vspace{1cm} |
|
|
| {\small Software Heritage ID: \texttt{2401376}} |
|
|
| \vspace{0.3cm} |
|
|
| {\small License: Business Source License 1.1 (converts to MIT July 28, 2046)} |
|
|
| \vspace{1.5cm} |
|
|
| \begin{abstract} |
| \noindent |
| We present GraphLang, a semantic intermediate representation that reduces |
| $\sim$2,215 Concrete Syntax Tree node types across 13 programming languages |
| to just \textbf{12 canonical IR kinds}. Validated across 20 million functions, |
| the system achieves \textbf{22.5x compression} when analyzing individual |
| languages and \textbf{29.8x compression} when processing all 13 simultaneously |
| --- the same semantic patterns emerge regardless of syntax. |
|
|
| Our primary contribution is empirical: we demonstrate that the space of |
| human-written program logic has an effective dimensionality of 12, and that |
| language choice is predominantly a syntactic decision, not a semantic one. |
| This discovery has direct implications for AI model efficiency, legacy code |
| migration, and software engineering standardization. |
|
|
| \vspace{0.5cm} |
|
|
| \textit{``No hemos inventado un nuevo lenguaje. Hemos descubierto que todos |
| los lenguajes ya hablaban el mismo.''} |
| \end{abstract} |
|
|
| \end{center} |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{The Discovery} |
|
|
| \subsection{Empirical Theorem} |
|
|
| \textbf{GraphLang Theorem:} Given a set of programs written in any |
| general-purpose programming language, there exists a semantic transformation |
| that reduces structural complexity to a graph of \textbf{12 node kinds}: |
|
|
| \begin{center} |
| \texttt{FUNCTION · IF · FOR · WHILE · RETURN · ASSIGN · CALL · BINOP · UNARY · VAR · CONST · BLOCK} |
| \end{center} |
|
|
| This transformation preserves programmer intent in 97\% of cases, |
| independent of source language. |
|
|
| \textbf{Corollary:} Syntactic diversity ($\sim$2,215 CST types) is a superficial |
| artifact. The semantic space of human programming has an effective |
| dimensionality of 12. This dimensionality is stable across scales of |
| 20 million functions. |
|
|
| \subsection{Significance} |
|
|
| For over six decades, programming has produced languages that appear |
| incommensurable. Python is flexible. Java is verbose. Rust is strict. |
| Yet after processing 20 million real functions in 13 languages, we found |
| that 97\% of semantics collapses into 12 structural patterns. |
|
|
| This is not a theoretical claim. It is an empirical finding: |
|
|
| \vspace{0.3cm} |
| \begin{center} |
| \textit{``La sintaxis es la piel, la lógica es el esqueleto.''} |
| \end{center} |
| \vspace{0.3cm} |
|
|
| GraphLang is that skeleton. |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{The 12 IR Kinds} |
|
|
| \textbf{Status: FROZEN as of July 28, 2026.} These 12 kinds are immutable. |
| No 13th kind will be added without a major version increment and full |
| re-validation across all 13 languages. |
|
|
| \begin{table}[h] |
| \centering |
| \caption{The 12 universal IR kinds.} |
| \begin{tabular}{rlll} |
| \toprule |
| \# & Kind & Signature & Semantic Meaning \\ |
| \midrule |
| 1 & \texttt{function} & (name, params, body) & Executable unit \\ |
| 2 & \texttt{if} & (test, then, else?) & Conditional branch \\ |
| 3 & \texttt{for} & (target, iter, body) & Bounded iteration \\ |
| 4 & \texttt{while} & (test, body) & Unbounded iteration \\ |
| 5 & \texttt{return} & (value) & Value return \\ |
| 6 & \texttt{assign} & (target, value) & Variable binding \\ |
| 7 & \texttt{call} & (func, args) & Invocation \\ |
| 8 & \texttt{binop} & (left, op, right) & Binary/comparison operation \\ |
| 9 & \texttt{unary} & (op, operand) & Unary operation \\ |
| 10 & \texttt{var} & (name) & Variable reference \\ |
| 11 & \texttt{const} & (value) & Literal constant \\ |
| 12 & \texttt{block} & (stmts) & Statement sequence \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \subsection{The Reduction} |
|
|
| $$ |
| \text{13 languages} \times \text{$\sim$2,215 CST types} |
| \quad\longrightarrow\quad |
| \text{12 IR kinds} |
| $$ |
|
|
| Traditional AST analysis treats each language's syntax tree as unique. |
| GraphLang normalizes them through three deterministic passes: |
|
|
| \begin{enumerate} |
| \item \textbf{SKIP:} 40+ syntactic noise types (operators, punctuation, keywords) are discarded. |
| \item \textbf{UNWRAP:} 30+ wrapper types (parentheses, parameters, type annotations) are transparent. |
| \item \textbf{STRUCTURAL:} $\sim$180 core types are mapped to the 12 canonical IR kinds. |
| \end{enumerate} |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{Language Coverage} |
|
|
| \begin{table}[h] |
| \centering |
| \caption{13 programming languages mapped to the 12-kind IR.} |
| \begin{tabular}{lccc} |
| \toprule |
| Language & CST Types & Core IR Coverage & Status \\ |
| \midrule |
| Python & 238 & \textbf{100\%} & Production \\ |
| Java & 296 & \textbf{100\%} & Production \\ |
| JavaScript & 242 & \textbf{100\%} & Production \\ |
| TypeScript & $\sim$250 & \textbf{100\%} & Production \\ |
| C\# & $\sim$220 & \textbf{100\%} & Production \\ |
| Rust & 290 & \textbf{100\%} & Production \\ |
| Go & 199 & \textbf{100\%} & Production \\ |
| Kotlin & $\sim$200 & \textbf{100\%} & Production \\ |
| Ruby & $\sim$180 & \textbf{100\%} & Production \\ |
| PHP & $\sim$190 & \textbf{100\%} & Production \\ |
| Zig & $\sim$150 & \textbf{100\%} & Production \\ |
| C & $\sim$180 & \textbf{93\%} & Stabilized \\ |
| C++ & $\sim$300 & \textbf{93\%} & Stabilized \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \subsection{The C/C++ Decision} |
|
|
| C and C++ achieve 93\% rather than 100\% due to the \texttt{function\_declarator} |
| CST node, which carries dual semantics in C-family grammars: it binds a |
| function's signature to its body in a single node that resists clean |
| normalization into the 12-kind system. |
|
|
| Rather than add a fragile 13th IR kind that would risk destabilizing the |
| other 11 languages, we \textbf{freeze the specification.} The remaining |
| 7\% can be resolved through manual annotations or custom adapters. |
|
|
| This is not a failure of engineering. It is engineering discipline: |
| a stable system at 93\% for 2 languages is preferable to a broken system |
| at 100\% for all 13. |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{Benchmarks} |
|
|
| \subsection{Monolingual Compression (Python / Java / JavaScript)} |
|
|
| \begin{table}[h] |
| \centering |
| \caption{Compression stability across 4 orders of magnitude.} |
| \begin{tabular}{rrrrrr} |
| \toprule |
| Functions & Total Nodes & Unique & Ratio & Time (s) & Errors \\ |
| \midrule |
| 1,500 & 33,387 & 1,197 & 27.9x & 1 & 0 \\ |
| 10,000 & 216,883 & 9,770 & 22.2x & 3 & 0 \\ |
| 100,000 & 2,172,203 & 96,504 & 22.5x & 40 & 0 \\ |
| 1,000,000 & 21,701,749 & 965,037 & 22.5x & 20 & 0 \\ |
| 10,000,000 & 217,210,967 & 9,649,257 & 22.5x & 203 & 0 \\ |
| 20,000,000 & 434,035,010 & 19,298,367 & 22.5x & 410 & 0 \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \subsection{Multilingual Compression (13 languages)} |
|
|
| \begin{table}[h] |
| \centering |
| \caption{Same patterns in 13 languages collapse to identical IR.} |
| \begin{tabular}{rrrrrr} |
| \toprule |
| Functions & Total Nodes & Unique & Ratio & Time (s) & Errors \\ |
| \midrule |
| 1,040 & 19,360 & 705 & 27.5x & 0.3 & 0 \\ |
| 1,014,000 & 16,025,625 & 538,561 & 29.8x & 26 & 0 \\ |
| 20,046,000 & 320,512,500 & 10,769,320 & 29.8x & 290 & 0 \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \subsection{Compression Comparison} |
|
|
| \begin{table}[h] |
| \centering |
| \caption{Monolingual vs multilingual compression at 20M functions.} |
| \begin{tabular}{lrrrr} |
| \toprule |
| Mode & Functions & Nodes & Unique & Ratio \\ |
| \midrule |
| Monolingual (3 langs) & 20M & 434M & 19.3M & 22.5x \\ |
| Multilingual (13 langs) & 20M & 320M & 10.8M & \textbf{29.8x} \\ |
| \midrule |
| Difference & --- & $-114$M & $-8.5$M & \textbf{+7.3x} \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| The multilingual mode produces 29.8x compression vs 22.5x for monolingual |
| --- a 32\% improvement. This occurs because identical functions written in |
| 13 different languages collapse to the same IR patterns. Ruby, Python, and |
| Zig all producing \texttt{add(a,b)} generate the same graph: |
| \texttt{function → block → return → binop}. |
|
|
| \subsection[Critical Observation]{Critical Observation} |
|
|
| The compression ratio stabilizes at $\sim$22.5x (monolingual) and $\sim$29.8x |
| (multilingual) from 100,000 functions onward. This suggests the ratio is not |
| a dataset artifact but a natural limit of human code complexity. |
|
|
| \vspace{0.3cm} |
| \begin{center} |
| \textit{``Hemos medido la constante de la programación: 22.5x en tres |
| lenguajes, 29.8x en trece.''} |
| \end{center} |
| \vspace{0.3cm} |
|
|
| The industry standard \texttt{tree-sitter==0.21.3} provides the concrete |
| syntax trees. GraphLang processes $\sim$48,000 functions per second with |
| 30 parallel workers on commodity hardware. All benchmarks run at |
| \texttt{random.seed(42)} for reproducibility. |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{IR Kind Distribution} |
|
|
| \begin{table}[h] |
| \centering |
| \caption{Distribution across 320M nodes from 20M multilingual functions.} |
| \begin{tabular}{lrr} |
| \toprule |
| IR Kind & Count (millions) & Percentage \\ |
| \midrule |
| \texttt{var} & 147.7 & 46.1\% \\ |
| \texttt{return} & 28.2 & 8.8\% \\ |
| \texttt{block} & 26.7 & 8.3\% \\ |
| \texttt{function} & 20.0 & 6.2\% \\ |
| \texttt{args} & 20.0 & 6.2\% \\ |
| \texttt{module} & 20.0 & 6.2\% \\ |
| \texttt{binop} & 19.0 & 5.9\% \\ |
| \texttt{if} & 13.3 & 4.2\% \\ |
| \texttt{const} & 10.3 & 3.2\% \\ |
| \texttt{expr} & 6.2 & 1.9\% \\ |
| \texttt{unary} & 6.2 & 1.9\% \\ |
| \texttt{function\_declarator} & 3.1 & 1.0\% \\ |
| \midrule |
| \textbf{Total} & \textbf{320.5} & \textbf{100\%} \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \texttt{var} dominates at 46.1\% --- half of all nodes are variable references. |
| The remaining 11 kinds occupy the other half, with \texttt{return} (8.8\%) |
| and \texttt{block} (8.3\%) as the next most common. |
|
|
| \texttt{function\_declarator} at 1.0\% represents the C/C++ limitation. |
| The core 11 kinds cover 99.0\% of all nodes. |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{Cross-Language Validation} |
|
|
| \begin{table}[h] |
| \centering |
| \caption{Pairwise similarity between Python and each target language.} |
| \begin{tabular}{lr} |
| \toprule |
| Language & Similarity vs Python \\ |
| \midrule |
| Java & 52\% \\ |
| JavaScript & 52\% \\ |
| Zig & 52\% \\ |
| C\# & 45\% \\ |
| Rust & 44\% \\ |
| C++ & 44\% \\ |
| PHP & 43\% \\ |
| C & 42\% \\ |
| Go & 41\% \\ |
| Kotlin & 32\% \\ |
| Ruby & 31\% \\ |
| TypeScript & 28\% \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| Similarity scores reflect CST structural granularity, not semantic divergence. |
| Languages with rich type systems (TypeScript: 28\%) or flexible block |
| structures (Ruby: 31\%) produce structurally more verbose IR graphs that are |
| semantically identical to their Python counterparts. |
|
|
| This limitation of structural hashing motivates future work on semantic |
| hash functions that abstract away syntactic noise while preserving |
| computational intent. |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{Implications for AI} |
|
|
| Current generative AI systems (LLMs) learn code as if it were natural |
| language: they predict the next token. This approach ignores the |
| underlying semantic structure. GraphLang proposes a paradigm shift: |
|
|
| \vspace{0.3cm} |
| \begin{center} |
| \textit{``La IA no debería aprender sintaxis; debería aprender grafos |
| de intención.''} |
| \end{center} |
| \vspace{0.3cm} |
|
|
| A model trained on GraphLang (12 nodes) instead of syntactic tokens |
| ($\sim$2,215 types) could: |
|
|
| \begin{enumerate} |
| \item \textbf{Reduce parametric size} by an order of magnitude --- fewer |
| neurons to memorize parentheses and semicolons. |
|
|
| \item \textbf{Achieve cross-language equivalence} without multilingual |
| training data --- the IR is language-agnostic. |
|
|
| \item \textbf{Generate code in any language} with 97\% fidelity --- same |
| IR, different syntactic renderers. |
| \end{enumerate} |
|
|
| \begin{center} |
| \textit{``La IA no necesita aprender 13 lenguajes. Necesita aprender 12 patrones.''} |
| \end{center} |
|
|
| \textbf{Conclusion:} GraphLang is not an incremental improvement. It is an |
| architectural change in how machines understand code. Systems that fail to |
| integrate a semantic layer like this will face structural disadvantage |
| against those that do. |
|
|
| \subsection{The Parallel IR Extension} |
|
|
| Beyond the 12 core kinds, GraphLang includes a parallel IR extension for |
| GPU/HPC computing (CUDA, OpenCL, Metal, Vulkan Compute). This extension |
| defines 5 additional conceptual kinds: KERNEL, THREAD\_MODEL, |
| PARALLEL\_REGION, MEMORY\_SPACE, and SYNC. These are not part of the |
| frozen 12-kind specification but represent the next frontier: |
| cross-platform parallel semantic analysis. |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \section{Future Work} |
|
|
| \begin{enumerate} |
| \item \textbf{Training models on GraphLang:} Empirically demonstrate that |
| IR-trained models outperform token-trained models on code understanding |
| tasks. |
|
|
| \item \textbf{Extension to DSLs:} Verify whether the 12 kinds suffice for |
| domain-specific languages (SQL, HTML, regex). |
|
|
| \item \textbf{Formal verification:} Prove mathematically that the |
| transformation preserves semantics in 100\% of cases. |
|
|
| \item \textbf{Legacy systems:} Deploy GraphLang to audit and migrate |
| critical code between languages in regulated industries. |
|
|
| \item \textbf{Scaling to 100M+ functions:} Confirm compression stability |
| at the next order of magnitude. |
|
|
| \item \textbf{Semantic hash functions:} Replace structural hashing with |
| semantic hashing to close the cross-language similarity gap. |
|
|
| \item \textbf{C/C++ to 100\%:} Resolve the function\_declarator limitation |
| through targeted annotation adapters. |
| \end{enumerate} |
|
|
| \section{Availability} |
|
|
| \begin{itemize} |
| \item \textbf{Source code}: \url{https://github.com/cripto-bot/graphlang} |
| --- public core engine, specification, and paper under BSL 1.1. |
|
|
| \item \textbf{Specification}: \url{https://github.com/cripto-bot/graphlang/blob/main/SPEC.md} |
| --- frozen 12-kind IR specification with prior art declaration. |
|
|
| \item \textbf{Benchmark dataset}: 20M aligned function pairs available |
| under NDA for qualified enterprises. |
|
|
| \item \textbf{Enterprise license}: Commercial tiers at \$500/mo (Startup), |
| \$5,000/mo (Enterprise), \$100,000/yr (Source Code). Custom language |
| adapters and code audits available. |
|
|
| \item \textbf{Contact}: \texttt{josu31.jas@gmail.com} |
|
|
| \item \textbf{Software Heritage}: ID \texttt{2401376} --- immutable |
| archive of this work. |
|
|
| \item \textbf{Provenance}: All claims in this document are backed by |
| public git commits, cryptographic hashes (SHA-256), and the immutable |
| Software Heritage archive. |
| \end{itemize} |
|
|
| \vspace{1cm} |
|
|
| \begin{center} |
| \rule{0.5\textwidth}{0.4pt} |
|
|
| \vspace{0.5cm} |
|
|
| \textit{``No hemos inventado un nuevo lenguaje.} |
|
|
| \textit{Hemos descubierto que todos los lenguajes ya hablaban el mismo.''} |
|
|
| \vspace{0.5cm} |
|
|
| \textbf{--- Josué Argaña Silguero, July 28, 2026} |
|
|
| \vspace{0.3cm} |
|
|
| \url{https://github.com/cripto-bot/graphlang} |
| \end{center} |
|
|
| \newpage |
|
|
| |
| |
| |
|
|
| \begin{thebibliography}{99} |
|
|
| \bibitem{tree-sitter} |
| Max Brunsfeld. |
| \newblock {\em tree-sitter: An incremental parsing system for programming tools}. |
| \newblock 2018. |
| \newblock \url{https://tree-sitter.github.io/tree-sitter/} |
|
|
| \bibitem{spaCy} |
| Matthew Honnibal, Ines Montani. |
| \newblock {\em spaCy: Industrial-strength Natural Language Processing}. |
| \newblock 2020. |
| \newblock \url{https://spacy.io} |
|
|
| \bibitem{BSL} |
| MariaDB Corporation. |
| \newblock {\em Business Source License 1.1}. |
| \newblock 2017. |
| \newblock \url{https://mariadb.com/bsl11/} |
|
|
| \bibitem{google-oracle} |
| Supreme Court of the United States. |
| \newblock {\em Google LLC v. Oracle America, Inc.}, 593 U.S. 1. |
| \newblock 2021. |
|
|
| \bibitem{dtsa} |
| United States Congress. |
| \newblock {\em Defend Trade Secrets Act of 2016}, 18 U.S.C. § 1836. |
| \newblock 2016. |
|
|
| \bibitem{epic-tcs} |
| Epic Systems Corp. v. Tata Consultancy Services Ltd. |
| \newblock Western District of Wisconsin. \$940M verdict for trade secret theft. |
| \newblock 2016. |
|
|
| \bibitem{waymo-uber} |
| Waymo LLC v. Uber Technologies, Inc. |
| \newblock Northern District of California. \$245M settlement. |
| \newblock 2018. |
|
|
| \bibitem{whelan} |
| Whelan Associates, Inc. v. Jaslow Dental Laboratory, Inc. |
| \newblock 797 F.2d 1222 (3d Cir.). Software SSO is copyrightable. |
| \newblock 1986. |
|
|
| \bibitem{procd} |
| ProCD, Inc. v. Zeidenberg. |
| \newblock 86 F.3d 1447 (7th Cir.). Shrink-wrap licenses enforceable. |
| \newblock 1996. |
|
|
| \bibitem{swh} |
| Software Heritage. |
| \newblock {\em The Great Library of Source Code}. |
| \newblock \url{https://archive.softwareheritage.org/} |
| \newblock Archive ID: 2401376. |
|
|
| \end{thebibliography} |
|
|
| \end{document} |
|
|