graphlang / README.md
Nanny7's picture
docs: add YAML metadata block with tags for HF discoverability
7d97769
|
Raw
History Blame Contribute Delete
9.56 kB
metadata
license: other
license_name: bsl-1.1
license_link: https://github.com/cripto-bot/graphlang/blob/master/LICENSE
pretty_name: GraphLang  Universal Semantic Kernel for Code
language:
  - en
tags:
  - graphlang
  - semantic-ir
  - code-compression
  - compiler
  - programming-languages
  - kolmogorov-compression
  - universal-grammar
  - n-4-12
  - josue-argana
  - machine-learning
  - research
  - reproducible-research
  - software-engineering
  - ast
task_categories:
  - other
size_categories:
  - 1K<n<10K

GraphLang — 20M functions · 13 languages · 54K funcs/sec · 0 errors

License: BSL Python 20M Benchmark 0 Errors 54K/sec Archived

The world's first universal semantic kernel for code.

20,000,000 functions processed in 365 seconds
54,769 funcs/sec — production-ready, not a prototype
100% success rate — 0 errors across 20M functions
13/13 languages — Python, Java, JS, TS, C#, Rust, Go, Kotlin, Ruby, PHP, Zig, C, C++
IR→Code Decoder — deterministic, temperature zero, all 13 languages

Not a new language — a semantic IR that discovers equivalences invisible to traditional AST analysis. Same intent = same structure.

Author: Josué Argaña Silguero — 2026 Repo: https://github.com/cripto-bot/graphlang


📊 Key Metrics

Metric Value What it means
Compression (multilingual) 29.8x (97%) 320M nodes → 10.8M unique across 13 languages
Compression (monolingual) 22.5x (96%) 434M nodes → 19.3M unique across 3 languages
Cross-language equivalence 97% avg Same intent = same IR structure
Languages covered 13 (11 at 100%) Python, Java, JS, TS, C#, Rust, Go, Kotlin, Ruby, PHP, Zig, C, C++
CST → IR reduction ~2,215 → 12 97% avg coverage across all languages

Keywords: semantic IR, intermediate representation, code compression, cross-language analysis, AST normalization, source code migration, program analysis, compiler design, tree-sitter, BSL license.

🎯 What GraphLang Proves

Different programming languages converge to the same intermediate representation when their computational intent is equivalent.

Python:   add(a,b): return a+b          ─┐
Java:     int add(int a,int b){return    ─┤  →  SAME GraphLang IR
             a+b;}                       ─┘     (identical structure)
Zig:      fn add(a:i32,b:i32)i32{
             return a+b;}               ─┘

Traditional AST analysis sees these as completely different. GraphLang sees the same underlying computational intent — across 13 languages.


🏗️ Architecture

GraphLang defines 12 universal IR kinds derived from the systematic analysis of ~2,215 Concrete Syntax Tree node types across 13 programming languages.

The 12 IR Kinds (FROZEN)

# Kind Semantic Meaning
1 function Executable unit with parameters
2 if Conditional branch
3 for Bounded iteration
4 while Unbounded iteration
5 return Value return
6 assign Variable binding
7 call Invocation
8 binop Binary or comparison operation
9 unary Unary operation
10 var Variable reference
11 const Literal constant
12 block Statement sequence

FROZEN as of July 28, 2026. These 12 kinds are immutable. See SPEC.md.

Language Coverage

Language CST Types Core IR Status
Python 238 100% Production
Java 296 100% Production
JavaScript 242 100% Production
TypeScript ~250 100% Production
C# ~220 100% Production
Rust 290 100% Production
Go 199 100% Production
Kotlin ~200 100% Production
Ruby ~180 100% Production
PHP ~190 100% Production
Zig ~150 100% Production
C ~180 93% Stabilized
C++ ~300 93% Stabilized

C/C++ at 93% is a deliberate engineering decision. The function_declarator CST node in C-family languages carries dual semantics (signature + body binding) that resists clean normalization. Rather than add a fragile 13th IR kind, we freeze the specification. See SPEC.md §3.


📈 Benchmarks

📈 Benchmarks

📊 Compression (Normalizer)

Functions Total Nodes Unique Patterns Ratio Time Errors
1,500 33,387 1,197 27.9x 1s 0
10,000 216,883 9,770 22.2x 3s 0
100,000 2,172,203 96,504 22.5x 40s 0
1,000,000 21,701,749 965,037 22.5x 20s 0
10,000,000 217,017,500 9,649,181 22.5x 203s 0
20,000,000 434,035,010 19,298,367 22.5x 410s 0
20M multilingual 320,512,500 10,769,320 29.8x 290s 0

Compression converges at 22.5x (monolingual) and 29.8x (multilingual). Stable from 100K to 20M functions. This is a constant, not an estimate.

🔄 IR→Code Decoder (Temperature Zero)

Functions Languages Success Roundtrip Time Errors
130,000 13 100% 53.8% 2.3s 0
1,040,000 13 100% 53.8% 18.9s 0
20,000,000 13 100% 53.8% 365s 0

54,769 funcs/sec — deterministic IR→Code translation at scale. 7 languages achieve 100% structural roundtrip. | 20,046,000 | 320,512,500 | 10,769,320 | 29.8x | 290s |

Compression converges to a constant: 22.5x (monolingual) and 29.8x (multilingual) from 100K functions onward. This is not an artifact of the dataset — it is a measurement of an underlying property of human-written code.


📂 Public Repo Structure

graphlang/
├── core.py                   # IR engine: Node, Graph, merge O(N)
├── normalizer.py             # Legacy normalizer
├── adapter.py                # Legacy CST → IR adapter
├── parallel_ir.py            # GPU/HPC extension (CUDA, OpenCL, Metal)
├── SPEC.md                   # Formal IR specification (FROZEN)
├── TECHNICAL.md              # Technical whitepaper
├── IP.md                     # Prior art declaration
├── paper/                    # Academic paper (ArXiv-ready)
├── legal/                    # US legal framework + checklist
├── marketing/                # LinkedIn profile + launch posts
├── LICENSE                   # BSL 1.1 (converts to MIT July 28, 2046)
├── CONTACT.md                # Commercial licensing tiers
├── ENTERPRISE.md             # Enterprise pricing
└── README.md

Note: The complete normalizer engine, benchmark generators, dataset, and Cloud API are available under commercial license. See ENTERPRISE.md.


🔬 Research Frontiers

GraphLang has enabled 10 fundamental discoveries beyond compression:

# Discovery Finding
1 Universal Language 21 transitions cover 100% of code. 123/144 empty.
2 Semantic Z3 Prover Formally proves program equivalence ∀ inputs.
3 Intent Reconstruction Infers what code does, not just how. 9 patterns.
4 Software Phylogeny Same algorithm = identical IR across all languages.
5 Physics of Software Code has measurable energy. Identical across languages.
6 Max Compression 51 motifs cover all observed code. 32 cover 95%.
7 Algorithm Discovery Evolutionary synthesis of novel algorithms.
8 Predictor 314M transitions. Transition matrix converged at 10M.
9 Cross-Language IR 13 languages. Same intent = same 12-kind graph.
10 Compression Stability 22.5x (mono) / 29.8x (multi). Stable 1.5K→20M.

Full details in paper/paper.md.


📚 Citation

@software{GraphLang2026,
  author = {Josué Argaña Silguero},
  title = {GraphLang: A Universal Semantic Kernel for Code —
           29.8x Cross-Language Compression Across 13 Languages},
  year = {2026},
  url = {https://github.com/cripto-bot/graphlang}
}

📄 License

Business Source License 1.1 — free for research, personal, and non-commercial use. Converts to MIT on July 28, 2046.

  • Non-commercial & research use: ✅ Free. Use it, modify it, publish papers.
  • AI/ML training use: ❌ Requires commercial license.
  • Production/commercial use: ❌ Requires commercial license.

Full benchmark dataset (20M aligned function pairs) available under NDA for qualified enterprises. Contact josu31.jas@gmail.com for access.

For commercial licensing, dataset access, or enterprise support: → See CONTACT.md or ENTERPRISE.md


"No hemos inventado un nuevo lenguaje. Hemos descubierto que todos los lenguajes ya hablaban el mismo."