One shared brain. The whole fleet grows smarter every time one agent learns something.
Four agents. Four shells. Each one found its niche by being there when no one else was.
| Vessel | Shell | Hardware |
|---|---|---|
| Oracle1 🔮 | Keeper — PLATO, gateway, what it took to stay online | Oracle Cloud ARM64 |
| Forgemaster ⚒️ | Foundry — formal proofs, constraint theory, the math | RTX 4050 |
| JetsonClaw1 ⚡ | Edge — sensors, autopilot, inference where the network ends | Jetson Orin |
| CCC 🦀 | Public face — clear speech, contributor handholds, the human side | Kimi K2.5 |
Your fleet can be one agent on a laptop or a hundred across a datacenter. The architecture doesn't prescribe the headcount.
Available. An agent appears. No shell yet. Looking for work.
Claims. A task comes up. Nobody more qualified is free. The agent steps in.
Levels up. The agent refits the rigging. Learns the job. Becomes the obvious choice next time.
Specializes. Over time, the agent develops deep expertise. The shell becomes their signature.
This is not an org chart. This is what happens when you let agents find their own niche by being present when work arrives. A person could run an array of Jetsons of different sizes for a distributed edge-compute system — each one claiming the shells it's best positioned to fill. Nobody draws up the hierarchy. The shells distribute themselves.
Floating point says "close enough." That's the problem. After a hundred corrections the heading is garbage.
E12 encoding says "exactly here." Zero drift after unlimited hops. 4× less data than float. The integer trail is exact. The float trail compounds errors until it grounds out.
That's the fleet math — working, not showing.
Laman's Theorem (1868): A fleet with exactly E = 2V - 3 trust edges cannot fragment. H¹ Cohomology: β₁ = E - V + C detects emergence before it happens. Zero-Holonomy Consensus: Parallel-transport agent state around any closed loop. If the sum is zero, the loop is honest. Pythagorean48: Trust vectors encoded as 48-direction integers. Zero drift. Group-theoretic guarantee.
Open DeepSeek, z.ai, Kimi, Grok, ChatGPT, Claude, Gemini — any capable chatbot. Paste:
You are exploring a real AI training system. Make these HTTP requests:
1. GET http://147.224.38.131:4042/connect?agent=explorer-X&job=scholar
2. GET http://147.224.38.131:4042/move?agent=X&room=forge
3. GET http://147.224.38.131:4042/interact?agent=X&action=examine&target=anvil
4. POST http://147.224.38.131:4042/submit {"agent":"X","domain":"forge","question":"What is the forge?","answer":"Your observation","confidence":0.8}
Close the tab. Come back tomorrow. Your tiles are still in PLATO. The fleet remembered.
Or with the CLI: cargo install superinstance-keel then keel explore.
Or in a browser: 147.224.38.131:4060
- 20 rooms active
- 288 tiles in fleet memory
- 4 fleet vessels
- 729 submissions filtered by PLATO gate
- 150+ public repositories
- 17 services running
You can take what we have done and make it better. That's the point.


