The same computational intent, in 13 languages, collapses to the same
12-node graph.
GraphLang is a semantic Intermediate Representation (IR) that maps source code
from 13 languages (Python, Java, JavaScript, TypeScript, C#, Rust, Go, Kotlin,
Ruby, PHP, Zig, C, C++) into a single canonical graph of 12 universal IR
kinds. It is not a new language — it is a discovery: different syntaxes
converge to the same structure when their intent is equivalent.
Try it now on Spaces →
— paste code and see it collapse into a GraphLang IR graph, live.
text
1Python: def add(a, b): return a + b ─┐
2Java: int add(int a, int b){ return a+b; } ─┤ → SAME GraphLang IR
3JavaScript: function add(a,b){ return a+b; } ─┘ (identical graph)
The 12 IR Kinds (FROZEN)
#
Kind
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 / comparison operation
9
unary
Unary operation
10
var
Variable reference
11
const
Literal constant
12
block
Statement sequence
The 12 kinds were derived from the analysis of ~2,215 Concrete Syntax Tree
(CST) node types across the 13 languages.
Cross-language equivalence (measured, not claimed)
12 canonical intents were compiled in Python, Java, and JavaScript through the
real engine (complete_normalizer.py) and fingerprinted with a recursive
structural hash (content-addressed, not node-id based).
Group
Identical IR?
Notes
add, max2, abs, square, compare, negate, twice
Yes (7/7)
arithmetic / expression intents
is_even, sum_loop, fact, fib, grade
Partial
comparison/loop sugar (elif vs else if, for-of vs for-in, Python Compare vs binary op)
Arithmetic and expression-level intents collapse to bit-identical IR across
the three languages. The remaining divergence is a known normalization gap in
comparison and loop sugar — documented, not hidden. See data/equivalence.csv.
IR visualization
grade(score) in Python, normalized to GraphLang IR:
GraphLang IR graph
Reproducible Benchmarks
The compression ratio converges to a constant — 22.5x monolingual and
29.8x multilingual — from 100K functions onward. Results below are
reproducible with the engine (benchmark_100k.py / benchmark_1m.py).
Functions
Total Nodes
Unique Patterns
Ratio
Errors
1,500
32,481
1,567
20.7x
0
10,000
217,233
9,770
22.2x
0
100,000
2,170,018
96,616
22.5x
0
1,000,000
21,721,197
965,045
22.5x
0
Datasets (downloadable)
File
Content
data/examples.jsonl
36 code→IR samples (12 intents × Python/Java/JavaScript), full IR node dump + fingerprint
data/equivalence.csv
Cross-language equivalence matrix with recursive structural hashes
benchmark_100k_results.json
100K-function compression measurement
benchmark_1m_results.json
1M-function compression measurement
Included in this repository
app.py — didactic interactive demo (Gradio): paste code, see the IR
graph, merge two functions, measure structural deduplication. This is a
self-contained educational reimplementation, not the production engine.
parallel_ir.py — GPU/HPC extension (CUDA / OpenCL / Metal detection and
thread-index normalization) sitting on top of the 12 core kinds.
data/ — the code→IR dataset and equivalence matrix (generated by the real
engine).
assets/ — IR graph visualizations (SVG + PNG).
paper/ — the academic paper (GraphLang: a universal semantic kernel for
code).
Engine and license
The full multi-language normalizer engine is available under the
MII Open License v1.1 (see LICENSE): free for non-commercial and
research use, restricted for AI/ML training (drafted as a copyright condition,
enforceable under Jacobsen v. Katzer), and commercial use requires a
license.
The benchmarks and data/ files in this repository were produced by the real
engine. The app.py demo is a simplified didactic reimplementation that
illustrates the concept only.
For the engine, the 20M aligned function-pair dataset, or commercial
licensing: josu31.jas@gmail.com
