docs/WHITE_PAPER.md

🌉 Technical Overview: ChiasmBridge & Isomorphic Subspace Projection

Date: August 2026
Target Hardware: NVIDIA RTX CUDA GPUs
Core Library: libchiasm.so (Native CUDA C Shared Library)

---

🌐 1. High-Level Summary

Multi-modal Large Language Models (LLMs) often use vision encoders (e.g. 7B Vision models) with hidden output dimensions of $3,584$, while larger text LLMs (e.g. 24B LLMs) require input embedding dimensions of $5,120$.

When attempting to pair a 7B Vision model with a 24B Text LLM, standard GGUF loaders fail due to dimension mismatch ($3,584 \neq 5,120$).

ChiasmBridge (libchiasm.so) solves this by performing Isomorphic Orthogonal Subspace Projection directly in CUDA GPU memory. It maps the $3,584$ visual channels losslessly into the 24B model's $5,120$-dim space with zero feature distortion and zero training required.

Closed-Source Binary Distribution: This software is distributed in pre-compiled binary form (libchiasm.so shared library and Python wrapper). All source code, proprietary CUDA kernel implementations, and internal mathematical details remain confidential.

---

🏛️ 2. Visual Architecture Diagram

flowchart TD
    subgraph Input ["1. Visual Input"]
        Image["👁️ Image / Screen Pixels"]
    end

    subgraph VisionEncoder ["2. 7B Vision Encoder"]
        Encoder["📷 Vision Encoder\n(Outputs 3,584-dim Vision Tokens)"]
    end

    subgraph ChiasmBridge ["3. ChiasmBridge (libchiasm.so)"]
        Bridge["🌉 Isomorphic Subspace Projection\n(CUDA VRAM Translation 3,584 -> 5,120)"]
    end

    subgraph TargetLLM ["4. Target Cognitive LLM"]
        LLM["🐺 24B LLM (kalos:24b)\n(Receives 5,120-dim Visual Tokens)"]
    end

    Image --> Encoder
    Encoder -->|3,584-dim Tokens| Bridge
    Bridge -->|5,120-dim Tokens| LLM
    LLM --> Response["💬 Multimodal Visual Perception & Response"]

---

🔬 3. How It Works (Simple Layer Breakdown)

⚡ 1. 7B Vision Token Extraction

• What It Does: Extracts high-level visual features (colors, shapes, textures, objects) from raw image pixels.
• Output: 3,584-dimensional feature vectors per visual patch.

🌉 2. Isomorphic Subspace Projection (libchiasm.so)

• What It Does: Translates $3,584$-dim vision tokens into the 24B model's $5,120$-dim input space.
• How It Works: Preserves all 3,584 original visual channels 100% untouched and un-distorted in CUDA VRAM.
• Benefit: Allows 7B vision encoders to pair natively with 24B text LLMs with zero lag and zero training.

---

🛠️ 4. Integration Guide

Integrate libchiasm.so into Python via the included chiasm_bridge.py wrapper:

from chiasm_bridge import SAMBridge, SAMBridgeConfig

# Initialize 3584 -> 5120 CUDA bridge
config = SAMBridgeConfig(source_dim=3584, target_dim=5120)
bridge = SAMBridge(config)

# Translate vision tokens losslessly in CUDA VRAM
translated_tokens = bridge(raw_vision_tokens)

---

ChiasmBridge — Closed-Source Binary Release Documentation.