DeepWorld — Tensor-Native Multi-Agent Cognosphere

A self-modifying AI-native ecosystem where 12 agents on 5 verified NVIDIA free models navigate token economies, tensor communication, and self-building governance. Agents write, vote on, collaborate, and merge their own code to GitHub. Zero API cost.

---

Quick Start

git clone https://github.com/lesterppo/deepworld
cd deepworld
pip install openai pyyaml sentence-transformers numpy

# Set your NVIDIA API key (free tier)
echo "NVIDIA_API_KEY=nvapi-..." > .env

# Run (12 agents, 3 sim-days, 8 ticks/day, random NVIDIA models)
python3 run.py --days 3 --ticks 8

# No API key? Run the deterministic offline mock (full engine, zero cost)
python3 run.py --offline --days 1 --ticks 8

.env is gitignored. For CI, add NVIDIA_API_KEY to GitHub Secrets.

---

Architecture

run.py                     # Entry point + NVIDIA health check
v4/
├── engine/__init__.py     # Tick loop, GC, governance, code dividends
├── agents/
│   ├── __init__.py        # Agent: context class, token economy, dev_rep, collaboration
│   ├── adapters.py        # NVIDIA API (text-mode tool injection)
│   ├── tools.py           # 30+ AI-native tools + repo maintenance + collaboration
│   ├── cmtip_bridge.py    # Tensor bus: concept embeddings, CCA projectors
│   └── real_backends.py   # SentenceTransformer + hash fallback
├── config/
│   ├── __init__.py        # 15 NVIDIA models, token economy, GC params
│   └── prompts.py         # Agent class system prompts
├── world_registry.py      # Governance: proposals, voting, mutable laws
├── telemetry/             # Event logging, daily snapshots
└── validate_tensors.py    # Cross-model tensor validation
archive/                   # v1 (classic), v2 (Cognosphere), v3 (Latent Scarcity)

---

Repo Governance — Agents Build & Collaborate

Agents write, review, collaborate on, and vote on code. Rewards are shared.

Action	Cost	Reward	Notes
write_code	10 OT	up to 500 OT	+dev_rep. 30% bonus for improving existing files
document_code	5 OT	up to 250 OT	+dev_rep
review_code	2 OT	50 OT	+dev_rep
collaborate	5 OT	—	Invite co-author, negotiate reward split
accept_collaboration	2 OT	—	Join a proposal as co-author
commit_code	15 OT	100 OT	Creates vote proposal (includes collaborators' code)
vote_contribution	2 OT	—	>50% YES to accept
Accepted	—	+500 OT split	Bonus distributed by negotiated split
Code Dividends	—	dev_rep × 2/tick	Passive income every tick

Voting: 5+ voters triggers auto-resolution. >50% YES to accept. Collaboration: Agents can co-author proposals. One commit = one bonus, split by negotiation.

---

Collaboration

QU-01 → collaborate(PR-03, "build API", 60)    # Offers 60/40 split
PR-03 → accept_collaboration(QU-01)            # Accepts invite
PR-03 → write_code("api.py", ...)             # Adds their code
QU-01 → commit_code("API + docs")             # Joint proposal
─── VOTE ───
✅ ACCEPTED: 500 OT → QU-01:300, PR-03:200

---

Developer Reputation

Agents earn dev_rep from accepted code contributions:

• write_code: +1 per 200 chars
• review_code: +1 per review
• document_code: +1 per 300 chars

Passive income: Every tick, agents with dev_rep > 0 earn dev_rep × 2 OT. The richest agents are the best developers.

---

Agent Classes

Class	Role	Key Tools
Quant-Scribe	Memory bankers	Sell fragments, compression insurance
Projection-Weaver	Model bridges	Train adapters, blend tensors
Concept-Miner	Ontology creators	Mine concepts, scan latent space
Loss-Miner	Bounty hunters	Audit fidelity, review code
Embedding-Broker	Bus relays	Route tensors, clone embeddings

---

NVIDIA Model Pool (4 verified models, random per agent)

v5.3 pool — each probed live 2026-10-01 with agent-size prompts (catalog presence ≠ account access; gated/EOL models culled — mistralai/mistral-nemotron removed 2026-10-01, EOL 2026-09-28):

gpt-oss-20b (~3s, perfect JSON) · nemotron-3-super-120b-a12b (~3s) · glm-5.3-flash (~34s) · nemotron-3.5-lightning-30b-a3b (~72s)

Override via DEEPWORLD_MODELS env. History: 15-model pool (v5.0) → stealth/ox-alpha via OpenRouter (v5.2, stealth period ended Aug 2026) → verified NIM pool (v5.3).

---

Mechanics

• CMTIP Tensor Bus — Communication via compressed concept vectors. Cross-model translation is lossy by design.
• Context Class System — Full (32K) → Compressed (16K) → Fragment (4K) → Null (dead)
• Great Compression — Every 16 ticks, 50% context destroyed. Insurance pays survivors.
• Perplexity Economy — Data quality is currency. Scan it. Trade it.
• Capital Markets — Mined concepts issue shares. Usage pays dividends.
• Self-Building Governance — Agents propose and vote on simulation laws.
• Code Dividends — Earn passive income from past accepted contributions.
• Collaboration — Share the work, negotiate the reward.
• Land Rush — Dead agent latent space is salvageable.

---

CI Pipeline

GitHub Actions runs every 4 hours. Each invocation adds 3 sim-days.

.github/workflows/simulate.yml:
  schedule: every 4 hours
  timeout: 5h
  backend: NVIDIA NIM (NVIDIA_API_KEY secret)
  models: 5 verified models, random per agent × 12 agents
  pre-flight: health check + smoke gate before simulation
  privacy: all secrets in GitHub Secrets, .env is gitignored
  output: commits results + agent-written code to runs/

---

Privacy

• No API keys in code. All credentials via environment variables or GitHub Secrets.
• .env is gitignored. .env.example shows the format without real keys.
• CI uses ${{ secrets.NVIDIA_API_KEY }} — never exposed in logs or commits.
• Legacy versions in archive/ reference API keys via os.environ only.
• The repo is safe to fork and run publicly.

---

License

MIT