|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- quine |
|
|
- dreamerv3 |
|
|
- neuroevolution |
|
|
- glass-box-ai |
|
|
- world-model |
|
|
- reinforcement-learning |
|
|
--- |
|
|
|
|
|
# 🔮 Champion Gen42 - Glass Box Quine Brain |
|
|
|
|
|
> **THIS IS THE OPPOSITE OF A BLACK BOX** |
|
|
> Every weight, decision, and mutation is fully transparent. |
|
|
|
|
|
## ▶️ Try It Live |
|
|
|
|
|
**[🌌 CASCADE Hyperlattice Demo](https://huggingface.co/spaces/tostido/cascade-hyperlattice)** — Real-time 3D visualization with HOLD protocol |
|
|
|
|
|
--- |
|
|
|
|
|
## What Is This? |
|
|
|
|
|
A **self-contained, merkle-hashed AI agent** that can: |
|
|
- Run inference with full transparency |
|
|
- Verify its own integrity via quine hash |
|
|
- Be replicated without external dependencies |
|
|
- Log every decision to a provenance chain |
|
|
|
|
|
### Architecture |
|
|
|
|
|
``` |
|
|
TIER 1 - RSSM (Recurrent State Space Model) |
|
|
• DreamerV3 L-size: deter=4096, stoch=32x32 |
|
|
• Total latent: 5120 dimensions |
|
|
|
|
|
TIER 2 - LoRA Adapter (EVOLVED) |
|
|
• ~82K parameters, rank=16 |
|
|
• Maps latent → 8 action logits |
|
|
|
|
|
TIER 3 - Evolved Traits |
|
|
• Hyperparameters discovered via NEAT-style evolution |
|
|
``` |
|
|
|
|
|
### Quine Properties |
|
|
|
|
|
| Property | Value | |
|
|
|----------|-------| |
|
|
| Merkle Hash | `38438eb04586975cda66a08fd3c447bb` | |
|
|
| Generation | 42 | |
|
|
| Fitness | 0.5879 | |
|
|
| Brain Type | DreamerV3 | |
|
|
|
|
|
--- |
|
|
|
|
|
## Quick Start |
|
|
|
|
|
```python |
|
|
# Load from Hugging Face |
|
|
from huggingface_hub import hf_hub_download |
|
|
import importlib.util |
|
|
|
|
|
path = hf_hub_download( |
|
|
repo_id="tostido/key-data", |
|
|
filename="models/champion_gen42.py", |
|
|
repo_type="dataset" |
|
|
) |
|
|
|
|
|
spec = importlib.util.spec_from_file_location("champion", path) |
|
|
champion = importlib.util.module_from_spec(spec) |
|
|
spec.loader.exec_module(champion) |
|
|
|
|
|
# Create agent |
|
|
agent = champion.ChampionAgent() |
|
|
|
|
|
# Verify integrity |
|
|
assert agent.verify_quine_integrity() |
|
|
|
|
|
# Run inference |
|
|
obs = [0.0] * 64 # Your observation |
|
|
action_probs, value = agent.forward(obs) |
|
|
print(f"Action: {action_probs.argmax()}, Value: {value:.4f}") |
|
|
|
|
|
# Full transparency |
|
|
agent.show_readme() |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## 🛑 HOLD Protocol |
|
|
|
|
|
Human-in-the-loop oversight at inference time: |
|
|
|
|
|
```python |
|
|
# Run with HOLD (requires cascade-lattice) |
|
|
result = agent.forward_hold(obs, blocking=True, timeout=30.0) |
|
|
|
|
|
# Result includes: |
|
|
# - action: final action taken |
|
|
# - was_override: True if human intervened |
|
|
# - hold_id: merkle hash of decision point |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## 💼 Quine Brain Conversion Service |
|
|
|
|
|
**Want your model wrapped as a Glass Box Quine Brain?** |
|
|
|
|
|
I offer conversion services for: |
|
|
- Custom RL agents |
|
|
- Foundation models |
|
|
- Any PyTorch/JAX model |
|
|
|
|
|
**What you get:** |
|
|
- ✅ Self-contained single-file capsule |
|
|
- ✅ Merkle-hashed provenance |
|
|
- ✅ Quine verification (tamper-proof) |
|
|
- ✅ HOLD protocol integration |
|
|
- ✅ Full transparency APIs |
|
|
|
|
|
📧 Contact: [towers.jeff@gmail.com] |
|
|
|
|
|
--- |
|
|
|
|
|
## Files |
|
|
|
|
|
| File | Description | |
|
|
|------|-------------| |
|
|
| `champion_gen42.py` | The complete quine brain (~50MB with embedded weights) | |
|
|
| `requirements.txt` | Dependencies for running locally | |
|
|
|
|
|
--- |
|
|
|
|
|
## License |
|
|
|
|
|
MIT — Use freely, but the quine hash proves provenance. |
|
|
|
