README.md

# Cozet --- Native SYNAXIM Base Model

**Architecture**: SYNAXIM (Symbiotic Native Axim Inference Machine)
**Author**: GRRN Research
**License**: Proprietary
**Status**: Phase 1 --- Architecture Validation (130M parameters)

---

## What is Cozet?

Cozet is a language model built from scratch on the **SYNAXIM architecture** --- a non-transformer design that replaces self-attention with a persistent associative memory matrix (the Symbiotic Gate).

Cozet is **not** a converted transformer. It is born native. Every weight is trained from random initialization through the M-matrix paradigm. No KV cache. O(1) memory. Infinite context.

## Architecture: Symbiotic Gate

Standard transformers compute attention as:
```
output = softmax(Q @ K^T / sqrt(d)) @ V    # O(n^2) compute, O(n) KV cache
```

SYNAXIM computes attention as:
```
gate   = sigmoid(mean(Q * K))               # scalar routing
M_new  = gate * M + (1-gate) * outer(k, v)  # O(1) persistent memory update
output = q @ M_new                           # associative retrieval
```

The M-matrix accumulates context through gated outer product updates and never grows. Memory is O(D^2) fixed regardless of sequence length.

## Model Configurations

| Config | Parameters | D | Layers | Heads | Intermediate | Status |
|--------|-----------|------|--------|-------|-------------|--------|
| Cozet-Small | 130M | 1024 | 12 | 16/4 GQA | 4096 | Phase 1 |
| Cozet-Medium | 1.3B | 2048 | 24 | 16/4 GQA | 8192 | Phase 2 |
| Cozet-Large | 7.2B | 4096 | 32 | 32/8 GQA | 14336 | Phase 3 |

## Training

Native pretraining from randomly initialized weights on FineWeb-Edu / RedPajama using truncated BPTT through the M-matrix chain.

## Links

- **Architecture**: SYNAXIM --- https://github.com/GRRN-MAKER/SYNAXIM
- **Theory**: Odyssey Meta-Cognitive Framework
- **Inference Engine**: SYNAXIM `.symb` format (pure NumPy, zero dependencies)

---

(c) 2026 GRRN Research. All rights reserved.