SGM-Substrate

Geometric learning at 50 KB. Binary locking. NAND gates. Hebbian geometry.

What This Is

A research system exploring geometric primitives as alternatives to neural scaling. Three components, each independently validated:

1. Binary Locking -- Three lines of code that eliminate catastrophic forgetting. 1.0000x retention across 1000 sequential tasks. Zero gradient interference. Exactly.

2. Lexical Similarity Engine -- Scores 0.76 Spearman on STS-B at 400 KB. Compiled to Arduino Uno at 3.4 KB flash, scoring 0.71. Proves STS-B is ~90% lexical overlap.

3. Geometric Seed -- Hebbian manifold that learns from pure observation. No labels, no supervision. Geometry is the knowledge. Self-organizes clusters and meaning axes via PCA.

Validated Results

See VALIDATED_RESULTS.md for complete benchmark tables with reproduction commands.

Claim	Value	Reproduction
Locking retention (1000 tasks)	1.0000x	python tests/test_retention.py
Gradient interference	0.000000	python tests/test_interference.py
STS-B (lexical engine)	0.7601	python tests/test_stsb.py
STS-B (Arduino STEM3)	0.7089	python archive/uno_deploy/test_stem3.py
STS-B (stem overlap, 0 params)	0.6854	python tests/test_stsb.py
NAND truth table	4/4	python tests/test_nand.py
GPU inference	1.1B pairs/sec	python test_bootstrap_inference.py

Honesty Note

The 0.76 STS-B score comes from the lexical n-gram system, not the geometric core. The geometric Seed scores 0.13 cold and -0.18 after training. This is a known limitation: Hebbian co-occurrence does not capture the surface-level similarity STS-B rewards. The Seed's value is in autonomous structure discovery, not benchmark optimization.

Structure

core/           Seed (Hebbian geometry), Mind (Markov + PCA + K-means), Gate Encoder
inference/      UnifiedEngine, ScalableEngine, HybridEngine, GeometricMind
training/       TrueLearner, ContrastiveLearner, SelfSupervised, Curriculum
microcode/      NAND gate mesh (5 bytes/gate, evolvable Boolean circuits)
ngram/          Lexical similarity engine (character n-grams, stem overlap)
hardware/       Arduino Uno deployment, GPU kernels
tests/          Retention, interference, STS-B, NAND, plasticity tests
validation/     Benchmark scripts
archive/        Development iterations, Uno deploy, experiment history
data/           Training corpus and targets

Quick Start

git clone https://github.com/ACD421/sgm-substrate.git
cd sgm-substrate
pip install -r requirements.txt
python tests/test_retention.py      # Binary locking proof
python tests/test_stsb.py           # STS-B benchmarks
python tests/test_nand.py           # Gate verification
python main.py                      # Full system demo

Dependencies

numpy>=1.24.0
scipy>=1.10.0
cupy-cuda12x>=13.0.0    # GPU acceleration
torch>=2.0.0            # For MNIST/CIFAR experiments
datasets>=2.14.0        # For STS-B benchmark

Author

Andrew Dorman -- Independent AI researcher, Southlake TX

• GitHub: ACD421
• Research: Geometric primitives for intelligence without neural scaling

License

Proprietary. See LICENSE for terms.