Harmonic Interface Engine

A working prototype of new AI technology. Not a model. Not a wrapper. Not a user interface. A geometric substrate that processes information through harmonic field dynamics across fabricated silicon-aligned cores.

THIS IS NOT CONVENTIONAL AI

RELEASE — FEBRUARY 2026

Harmonic Engine V2

1,045 lines · 27 MB substrate · 702 prompts/sec · No GPU required

DOWNLOAD PACKAGE GITHUB

What This Is

This is a working prototype of an entirely new AI technology. It is an interface engine — a geometric substrate that replaces training with fabrication. Rather than learning weights through iterative gradient descent, the engine encodes geometric relationships directly into a tetrahedral lattice structure aligned with the cubic diamond crystal geometry (Fd3m space group) of the silicon it runs on.

The result is a system that achieves its operational throughput in 27 MB of RAM. Not gigabytes. Not terabytes. By eliminating the dead weight that conventional architectures carry as trained parameters, the engine operates at a fraction of the resource cost while maintaining geometric fidelity.

Measured Performance

27 MB

SUBSTRATE MEMORY

702

PROMPTS / SEC

1.4 ms

END-TO-END LATENCY

383K

TOKENS / SEC

370K

VOCABULARY

0 GPU

REQUIRED

Self-optimized on DGX (20 CPU, 128 GB RAM). Benchmark suite included in the release.

Architecture

Signal → [Junction Sensors] → [Harmonic Field] → [Geometric Cores] → Response

9 sensor types per core · shared lock-free field · 200 parallel domain-routed cores

Geometric Cores. Each core carries a unique 12-dimensional identity fingerprint derived from its own sensor array geometry. Nine sensor types detect signal characteristics through kernel convolutions. Six domain buffers process information from different geometric angles. Core identity is fabricated, not trained — it cannot drift.

Harmonic Field. The shared field is how cores perceive each other. Every core writes its output signature; every core reads the composite interference pattern. Coordination emerges from geometry, not from message passing or routing tables. The V2 implementation uses lock-free sharded composites with lazy evaluation for 71% throughput gain over V1.

Parallel Dispatch. OS-level shared memory enables true multiprocess parallelism. The field array lives in shared memory with zero-copy numpy access from all worker processes. At 128 cores this delivers 3.5x throughput over single-process execution.

Self-Optimizer. Measures actual throughput at multiple core counts and selects the configuration achieving 90% of peak. The throughput curve saturates early — the optimizer finds the plateau and avoids wasted resources beyond it.

How It Differs

	Conventional Model	Harmonic Interface Engine
Architecture	Transformer layers	Geometric lattice cores
Knowledge	Trained weights	Fabricated geometry
Memory	4–140 GB	27 MB
Learning	Gradient descent	One-pass geometric encoding
Scaling	More parameters	More cores (shared field)
Substrate	Abstract computation	Silicon-aligned (Fd3m)

Package Contents

harmonic_v1.pyMain engine — CLI, HTTP API, benchmark (1,045 lines)
fused_harmonic_substrate.pyCore substrate — 200 cores, harmonic field V2 (1,327 lines)
parallel_dispatch.pyShared memory multiprocess dispatcher (519 lines)
self_optimizer.pyThroughput-based core count optimization (219 lines)
geometric_codebook.pyGeometric decoder and grid encoder (1,002 lines)
codebook_expansion.pyDynamic codebook learning (954 lines)
governance_lattice.pyCouncil governance system (753 lines)
translation_table.json370K word geometric translation table
codebook_learned.jsonLearned geometric essence patterns

Usage

# Install
tar xzf harmonic_engine_v1.tar.gz
cd harmonic_engine_v1

# Interactive CLI (autoscale cores)
python3 harmonic_v1.py

# Self-optimize core count for this system
python3 harmonic_v1.py --cores auto

# HTTP server (Ollama-compatible API)
python3 harmonic_v1.py --http --port 11434

# Benchmark
python3 harmonic_v1.py --benchmark

Requirements: Python 3.10+ and NumPy. No GPU. No model downloads. No API keys.