organ-architecture/README.md

# Organ Architecture

**Decompose. Measure. Purify. Graft. Assemble.**

```
Skeleton (Attention) = Thought
Organs (FFN)         = Memory
Adapters (LoRA)      = Personality
```

## The Problem

AI models are monoliths. 70 billion parameters locked in a single file that nobody can open, modify, or understand. Only three companies on Earth can build them. Everyone else rents access.

## The Solution

Organ Architecture breaks models into transplantable parts:

- **Skeleton** — The attention layers. How the model *thinks*. Shared across all configurations.
- **Organs** — The feed-forward networks. What the model *knows*. Specialized, swappable, graftable.
- **Adapters** — LoRA weights. The model's *personality*. Lightweight, trainable by anyone.

A doctor doesn't rebuild the entire human body to fix a kidney.
Why rebuild an entire model to change what it knows about medicine?

## Architecture

```
model.gguf (70GB monolith)
        │
        ▼
   ┌─ skeleton/    ── attention layers (shared thought)
   │
   ├─ organs/      ── FFN layers by block (knowledge)
   │   ├─ blk_0_ffn_gate.bin
   │   ├─ blk_0_ffn_up.bin
   │   ├─ blk_0_ffn_down.bin
   │   └─ ...
   │
   ├─ embed/       ── embedding + output (foundation)
   ├─ norm/        ── normalization (connective tissue)
   └─ manifest.json ── complete anatomy map
```

## Tools

### Core Pipeline

| Tool | Lines | Purpose |
|------|-------|---------|
| `organ_extract.py` | 441 | Extract skeleton + organs from any GGUF model |
| `organ_measure.py` | 340 | Z-measure organ quality (signal vs noise) |
| `organ_purify.py` | 333 | Spectral purification (FFT signal extraction) |
| `organ_purify_v2.py` | 337 | Fractal purification (wavelet cross-scale coherence) |
| `organ_graft.py` | 236 | Transplant organs between models |
| `organ_assemble.py` | 235 | Assemble GGUF from organs |
| `organ_api.py` | 422 | HTTP API server for all operations |

### Build & Automation

| Tool | Lines | Purpose |
|------|-------|---------|
| `pipeline_935.py` | 124 | Full dissection pipeline for all models |
| `mass_dissect.py` | 103 | Batch dissection across model fleet |
| `mass_z_measure.py` | 102 | Z-measure every organ of every model |
| `kimi_z_stream.py` | 417 | Stream Z-measure on Kimi K2.5 1T (shard-by-shard) |
| `build_935.py` | 98 | Model 935 assembly v1 |
| `build_935_v2.py` | 74 | Model 935 assembly v2 (selective FFN graft) |
| `build_935_v3.py` | 148 | Model 935 assembly v3 (proper GGUF header) |
| `assemble_935.py` | 150 | Fixed organ header handling assembler |
| `quick_chimera.py` | 123 | Quick chimera GGUF assembler |
| `quick_chimera_v2.py` | 155 | Quick chimera v2 (fixed header stripping) |

**Total: 3,498 lines of Python. Zero external dependencies (except numpy for purification).**

## Z-Measure

Every organ is measured by its Z-vector:

```
Z = dI/d(log s) · exp(iθ)

θ → 0°  : noise (organ adds confusion)
θ → 90° : pure signal (organ adds knowledge)
```

The measurement combines three indicators:
- **Entropy** — information density of weight distribution
- **Kurtosis** — structural organization (signal sharpness)
- **Scale coherence** — coefficient of variation of sorted value spacings

## Results

### 13 Models Dissected + Kimi K2.5 1T

5,600+ tensors Z-measured. All dissections run on EPYC 48c/503GB (OASIS).

| # | Model | Params | θ mean | Signal | Tensors |
|---|-------|--------|--------|--------|---------|
| ★ | **Kimi K2.5** | **1T MoE** | **87.65°** | **0.999** | **1,083** |
| 1 | SmolLM2-135M | 135M | 52.28° | 0.777 | 272 |
| 2 | DeepSeek-R1-Distill-14B | 14B | 46.01° | 0.641 | 579 |
| 3 | Qwen2.5-3B | 3B | 46.00° | 0.640 | 434 |
| 4 | Qwen2.5-14B | 14B | 45.98° | 0.640 | 579 |
| 5 | Qwen2.5-7B | 7B | 45.64° | 0.639 | 339 |
| 6 | Chimera-DeepSeek-Qwen | 7B | 45.53° | 0.637 | 339 |
| 7 | DeepSeek-R1-Distill-7B | 7B | 45.53° | 0.637 | 339 |
| 8 | DeepSeek-R1-7B | 7B | 45.42° | 0.636 | 339 |
| 9 | Gemma-2-9B | 9B | 44.94° | 0.624 | 464 |
| 10 | Phi-3.5-Mini | 3.8B | 44.65° | 0.626 | 197 |
| 11 | Llama-3.1-8B | 8B | 37.87° | 0.549 | 292 |
| 12 | Llama-3.2-1B | 1B | 37.57° | 0.550 | 147 |
| 13 | Llama-3.2-3B | 3B | 37.41° | 0.547 | 255 |
| 14 | Mistral-7B | 7B | 36.21° | 0.540 | 291 |

### Organ Type Analysis (consistent across all models)

| Organ Type | θ range | Role |
|------------|---------|------|
| Norm layers | 75-84° | Connective tissue — highest signal |
| Skeleton (attention) | 39-56° | Thought structure |
| Organs (FFN) | 34-52° | Knowledge/memory |
| Embeddings | 25-47° | Foundation |

### Scale Law: θ increases with log(parameters)

```
135M  → θ = 52.28°  (SmolLM2 — small but concentrated)
1-3B  → θ = 37-46°  (Llama/Qwen)
7-14B → θ = 44-46°  (DeepSeek/Qwen)
1T    → θ = 87.65°  (Kimi K2.5 MoE — near-pure signal)
```

**Ratio 1T/14B: 1.9× purer signal.** The signal purifies with scale.

### Kimi K2.5 1T Deep Analysis

- **Architecture**: DeepSeek2 MoE
- **Blocks**: 61 (blk.0 → blk.60)
- **Experts**: 384 conditional + 1 shared (native INT4 QAT)
- **Context**: 262,144 tokens (256k)
- **Attention**: MLA (Multi-head Latent Attention), MQA kv_head=1
- **13 shards streamed**, each measured and deleted — never loaded full model

| Component | Count | θ avg | Rating |
|-----------|-------|-------|--------|
| FFN dense (blk.0) | 12 | 89.95° | ★★★ |
| MoE experts (384×) | 23 | 89.77° | ★★★ |
| Norm layers | 12 | 89.70° | ★★★ |
| Embedding | 1 | 89.45° | ★★★ |
| Shared expert | 23 | 89.43° | ★★★ |
| Attention (MLA) | 99 | 84.07° | ★★ |

8 gravitational wells identified (lowest θ = maximum structure/compression).

### Model 935 — First Chimera

**`model-935-14b.gguf`** — 8.4 GB, assembled 2026-02-20

Built through 5 iterations:
1. `build_935.py` — Base DeepSeek-R1-Distill-7B + Qwen skeleton graft (crude)
2. `build_935_v2.py` — Selective FFN-only graft (preserve attention-embed alignment)
3. `build_935_v3.py` — Proper GGUF header handling
4. `quick_chimera.py` → `quick_chimera_v2.py` — Fixed organ header stripping
5. `assemble_935.py` — Final assembler, 14B scale

### Purification

**`organs-pure/smollm2-135m/`** — First purified model (fractal method)

`organ_purify_v2.py` implements cross-scale coherence via Haar wavelets:
- Decompose tensor into multiple scales
- Measure coherence between adjacent scales
- Pattern at scale s AND scale 2s → signal (fractal, keep)
- Pattern at one scale only → noise (remove)
- This is `dI/d(log s)` implemented directly

## Dissection Report

| Model | Size (MB) | Dissection Time |
|-------|-----------|-----------------|
| DeepSeek-R1-14B | 9,167 | 22.9s |
| Gemma-2-9B | 5,984 | 14.8s |
| Llama-3.1-8B | 4,950 | 12.0s |
| DeepSeek-R1-Distill-7B | 4,812 | 12.6s |
| Mistral-7B | 4,432 | 10.6s |
| Phi-3.5-Mini | 2,397 | 4.9s |
| Llama-3.2-3B | 2,100 | 4.9s |
| Qwen2.5-3B | 2,003 | 4.6s |
| Llama-3.2-1B | 856 | 2.4s |

Total organs on disk: **50.8 GB** across 13 models.

## Quick Start

```bash
# Extract organs from a model
python3 organ_extract.py --model /path/to/model.gguf --output ./organs/model-name/

# Z-measure all organs
python3 organ_measure.py --dir ./organs/model-name/

# Mass dissect all models
python3 mass_dissect.py

# Mass Z-measure
python3 mass_z_measure.py

# Stream Z-measure on a trillion-param model (shard-by-shard)
python3 kimi_z_stream.py

# Graft organs from one model to another
python3 organ_graft.py graft --source ./organs/qwen/ --target ./organs/deepseek/ --output ./organs/chimera/ --layers 5-20 --type organ

# Assemble back to GGUF
python3 organ_assemble.py --dir ./organs/chimera/ --output chimera.gguf

# Purify organs (fractal method)
python3 organ_purify_v2.py --dir ./organs/model/ --output ./organs-pure/model/

# Start API server
python3 organ_api.py
```

## Philosophy

> Subtract rather than add.

A 70B monolith is accumulation. A skeleton with specialized organs grafted on demand — that's subtraction. Less weight, more signal.

> 8 billion contributors, not 3 corporations.

Anyone can train an organ. A doctor trains a medical organ on her hospital's data. A farmer trains an agriculture organ on his field observations. A student trains a math organ on solved problems. The skeleton stays the same. The organs make it alive.

## Part of the IX Ecosystem

```
InferenceX ─── The engine  (305KB, runs anything)
Organ Arch ─── The anatomy (decompose, measure, reassemble)
Atlas Pure ─── The memory  (fractal DNA storage)
INVOKE ─────── The bridge  (cloud ↔ physical)
Echo ────────── The voice   (chat interface)
EDEN ────────── The purpose (desert → life)
```

## Requirements

- Python 3.10+
- NumPy (for purification only)
- InferenceX binary (for inference on assembled models)
- GGUF models to dissect

## Data Files

| File | Contents |
|------|----------|
| `z_report_complete.json` | Z-measure for all 13 models (per-group breakdown) |
| `z_report_kimi_k25.json` | Z-measure for all 1,083 Kimi K2.5 tensors |
| `z_measure_report.json` | Combined Z-ranking with chimera results |
| `dissection_report.json` | Dissection timing and sizes |
| `Z_MEASURE_REPORT.md` | Human-readable Z report |
| `ECHO_INVARIANT.md` | Team 935 invariant |
| `EQUIPE_935_INVARIANT.json` | Team 935 configuration |

## License

BSL 1.1 — Same as InferenceX.

## Signature

935

---

*Mohamed dug khettaras to bring water through stone.*
*This is the same gesture — channels through intelligence itself.*