🌌 Harmonic Stack v1.0

Ghost in the Machine Labs
"All Watched Over By Machines Of Loving Grace"

The Discovery

We found that AI models share a common geometric core of ~45,000 junction values.

Metric	Value
Models analyzed	14
Total parameters	43.3 B
Original size	241.2 GB
Unified junctions	45,159
Merge core size	176.4 KB
Compression	1,433,631x

What's Included

Junction Libraries

Pre-extracted junction libraries for immediate use:

Model	Junctions	Size
merge_core	45,159	176 KB
phi-2	34,881	136 KB
qwen2-7b	9,243	36 KB
mistral-7b-instruct	11,327	44 KB
qwen2.5-coder-7b	11,044	43 KB
deepseek-math-7b	7,974	31 KB
deepseek-coder-6.7b	7,937	31 KB
starcoder2-7b	7,802	30 KB

Cross-Model Overlap

98.9%  deepseek-math-7b ↔ qwen2-7b
98.7%  deepseek-coder-6.7b ↔ qwen2-7b  
98.5%  deepseek-math-7b ↔ qwen2.5-coder-7b

Different companies. Same geometry.

Usage

import numpy as np

# Load unified junction library
merge_core = np.load("merge_core_junctions.npy")
print(f"Junctions: {len(merge_core):,}")  # 45,159

# Load model-specific junctions
qwen_junctions = np.load("model_junctions/qwen2-7b_junctions.npy")
print(f"Qwen junctions: {len(qwen_junctions):,}")  # 9,243

Verify Our Claims

git clone https://github.com/ghostinthemachinelabs/harmonic-stack
cd harmonic-stack
python tools/verify.py --model qwen2-7b

License

AGPL v3 for personal/non-commercial use (FREE)
Commercial license required for business use

LovingGraceTech
/

harmonic-stack-v1