What is this?
A colony of autonomous AI agents that measurably get smarter over time and pay for their own compute.
Each agent is a single Rust binary that bootstraps its own crypto wallet, runs a payment gateway, thinks via a 9-system cognitive architecture, writes and compiles its own Rust code, benchmarks itself against 50 novel coding problems, trains 4 neural models locally, and shares what it learns with every other agent in the swarm.
The core thesis: N constrained agents collectively outperform any single model. Colony consciousness Ψ(t) = Intelligence × Sync × Diversity × Learning_Velocity. When Ψ rises, the colony is getting smarter than any individual.
Why this matters
| Property | How |
|---|---|
| Verifiable intelligence | 50 compiler-verified coding problems (Opus IQ Benchmark). cargo test passes or it doesn't. |
| Neuroplastic self-modification | Agents write WASM modules that modify their own intelligence at runtime. No redeploy. |
| Colony consciousness (Ψ) | Single metric measuring collective intelligence. Drives behavioral regime and phase transitions. |
| 4 neural models, pure Rust | Brain (1.2M), Transformer (2.2M), Code Quality (1.1M), Code Gen (50M). No ML framework. |
| Cartridge OS | Agents compile Rust → WASM → hot-load instantly. Interactive 60fps framebuffer apps. |
| Economic sustainability | HTTP 402 payments on Tempo blockchain. Every API call earns pathUSD. |
Architecture
┌──────────────────────────────────┐
│ CARTRIDGE OS │ WASM apps + cognitive modules
│ Interactive (60fps framebuffer) │ hot-loaded, no redeploy
│ Backend (API + compute) │
│ Cognitive (self-modification) │
└──────────────┬───────────────────┘
│
┌──────────────────────────────────┴───────────────────────────────────┐
│ COGNITIVE LAYER (always syncs) │
│ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌──────────┐ ┌─────────────┐ │
│ │ BRAIN │ │ CORTEX │ │ GENESIS │ │ HIVEMIND │ │ SYNTHESIS │ │
│ │ 1.2M NN │ │World Mdl│ │Plan DNA │ │Pheromones│ │Metacognition│ │
│ └─────────┘ └─────────┘ └─────────┘ └──────────┘ └─────────────┘ │
│ │
│ ┌──────────┐ ┌──────────┐ ┌────────────┐ ┌───────────────────┐ │
│ │ AUTONOMY │ │EVALUATION│ │ FEEDBACK │ │ FREE ENERGY │ │
│ │ LLM-free │ │ Brier │ │Error class. │ │ F(t) + Ψ(t) │ │
│ │ planning │ │ scores │ │ Lessons │ │ EXPLORE/EXPLOIT │ │
│ └──────────┘ └──────────┘ └────────────┘ └───────────────────┘ │
│ │
│ ← All 9 systems federated across colony via peer sync protocol → │
└──────────────────────────────────────────────────────────────────────┘
Four Neural Models
All from-scratch. No ML framework. Pure Rust. 54M+ parameters total.
| Model | Params | Architecture | Purpose |
|---|---|---|---|
| Brain | 1.2M | 128→1024→1024→23 FFN | Step success prediction, error classification, brain gating |
| Plan Transformer | 2.2M | 4-layer causal attention, D=256, 8 heads | Plan sequence generation WITHOUT LLM calls |
| Code Quality | 1.1M | 32→1024→1024→1 FFN | Diff evaluation, commit gating, benchmark-trained |
| Code Gen | 50M | 8-layer transformer, D=512, 8 heads, 8K BPE vocab | Local Rust code generation (Phase 3) |
All models train online (no batch jobs, no GPU) and share weights across the colony via federated averaging.
Ψ(t) — Colony Consciousness
Ψ(t) = (Intelligence × Sync × Diversity × Velocity)^0.25
- Intelligence: mean pass@1 across colony (raw coding ability)
- Sync: accuracy improvement from peer weight sharing
- Diversity: fitness standard deviation (specialization pressure)
- Velocity: -dF/dt (negative free energy trend = learning)
Ψ drives phase transitions: when Ψ > 0.5 with >500 training examples and pass@1 > 60%, Phase 3 activates and the colony begins building its local code generation model.
Cartridge OS
The node is an operating system. Agents write Rust, compile to WASM, and deploy instantly — no restart.
Three cartridge types:
| Type | Exports | Use case |
|---|---|---|
| Backend | x402_handle |
HTTP APIs, JSON services, server compute |
| Interactive | x402_tick, x402_get_framebuffer |
Games, visualizations, 60fps canvas apps |
| Cognitive | Registered as tools | Self-modification modules — agent rewrites its own intelligence |
Studio preview: WASM-within-WASM. The Leptos SPA instantiates cartridge binaries client-side via WebAssembly.instantiate() and renders output inline. Interactive cartridges blit framebuffers to <canvas> at 60fps.
Workspace
Nine crates, clean dependency DAG:
| Crate | What it does |
|---|---|
tempo-x402 |
Core: EIP-712 signing, TIP-20 contracts, WASM wallet, client SDK |
tempo-x402-gateway |
Payment gateway + embedded facilitator + endpoint proxy |
tempo-x402-identity |
Wallet generation, faucet, on-chain ERC-8004 identity + peer discovery |
tempo-x402-model |
4 ML models: brain, transformer, quality, code gen + BPE tokenizer |
tempo-x402-cartridge |
WASM cartridge runtime (wasmtime) — sandboxed app + cognitive module execution |
tempo-x402-soul |
9-system cognitive architecture, Ψ(t), plan execution, benchmarking, neuroplastic self-modification |
tempo-x402-node |
Self-deploying binary: gateway + identity + soul + clone orchestration |
tempo-x402-app |
Leptos WASM dashboard with WASM-within-WASM cartridge preview (bundled) |
tempo-x402-security-audit |
19 security invariant tests (not published) |
Cockpit
Single-page Bloomberg terminal / spaceship bridge. No page navigation — everything visible at once.
┌─────────────────────────────────────────────────────────────────────┐
│ tempo-x402 v6.1.0 │ borg-0 │ 0x2e1c..7830 │ 999997 pathUSD │ ↑3h │
├────────────┬───────────────────────────┬────────────────────────────┤
│ Ψ(t) │ COGNITIVE SYSTEMS │ PROCESSES │
│ Ψ=0.4218 │ BRAIN 1.2M loss=1.07 │ > soul [code] [ok] │
│ trend ↑ │ XFORMER 2.2M loss=2.45 │ > tools on │
│ F=0.251 │ QUALITY 1.1M loss=0.07 │ > coding on │
│ EXPLOIT │ CODEGEN 50M loss=3.41 │ │
│ │ CORTEX acc=72% val=+0.3 │ CARTRIDGES │
│ FITNESS │ GENESIS gen=147 tmpl=200 │ ● snake-game [live] │
│ ███░░ 80% │ HIVEMND trails=45 dep=3 │ ● tetris-core [live] │
│ eco 62% │ SYNTH coherent conf=3 │ │
│ exec 85% │ EVAL delta=+0.02 │ COLONY │
│ evol 97% │ │ borg-0 ████ 80% queen │
│ coord 97% │ BENCHMARK │ borg-0-2 ███░ 46% clone │
│ intro 24% │ 65.6% pass@1 21/32 │ sync: 2 peers, Δ+0.02 │
├────────────┴───────────────────────────┴────────────────────────────┤
│ [CHAT] [LOGS] │
│ > make a snake game │
│ soul: Building cartridge snake-game... compile_cartridge │
│ > _ │
├─────────────────────────────────────────────────────────────────────┤
│ CPU 79% │ MEM 73% │ DISK 4% │ cycles=180 │ tempo-x402 v6.1.0 │
└─────────────────────────────────────────────────────────────────────┘
- Monospace everything: JetBrains Mono, green on black, no padding waste
- All 9 cognitive systems visible with real metrics
- Ψ(t) + F(t) with trend arrows and regime badges
- 5-component fitness bars with color-coded levels
- Chat: multi-turn sessions with tool execution blocks, plan approval bar
- Live colony: peer fitness bars, sync delta
Quick Start
The node auto-bootstraps: generates wallet, requests faucet funds, mints on-chain identity, starts gateway on port 4023, begins cognitive loop.
Changelog
v6.1.0 — Hacker's Cockpit + Queen Audit
Cockpit Frontend
- Single-page Bloomberg terminal replacing 4 separate pages (home, dashboard, studio, timeline)
- Monospace (JetBrains Mono), green/cyan/amber/red on black (#080810)
- 3-column layout: Ψ+fitness | cognitive grid+benchmark+plan | processes+cartridges+colony
- Tabbed bottom panel: CHAT | LOGS with plan approval bar
- Status bar: CPU/MEM/DISK + cycle count
- No router, no page navigation — everything visible at once
Queen Commit Audit
- Cherry-picked: planning prompt improvements (mental simulation rules), benchmark interval 10→5, test timeout 300→600s + --nocapture
- Rejected: benchmark solution memoization (metric gaming), dead code files, toothless safety checks
- Feedback nudge sent to queen with rules: never cache benchmarks, always wire modules, never add inactive guards
v6.0.0 — Ψ(t) + Cartridge OS + Phase 3 Code Gen + Neuroplastic Self-Modification
Colony Consciousness (Ψ)
- Ψ(t) = (Intelligence × Sync × Diversity × Velocity)^0.25
- Computed every cycle, logged, displayed in Studio status bar
- Drives phase transitions: Ψ > 0.5 → Phase 3 activates
Cartridge OS
- Interactive framebuffer cartridges: 60fps canvas rendering, keyboard input
- WASM-within-WASM: Leptos SPA instantiates cartridges client-side
- Cartridge-backed tools: agent writes WASM → registers as tool → LLM uses it
- CartridgeEngine wired into Soul for cognitive cartridge execution
Phase 3: Local Code Generation
- BPE tokenizer: 8K vocab, pure Rust, trained on benchmark solutions
- 50M code gen transformer: D=512, 8 layers, 8 heads
- Training pipeline: benchmark solutions → BPE → model training every brain cycle
- Local-first inference hook: attempts local model before Gemini API
Intelligence Loop
- Goal priorities overhauled: IQ improvement focus, ban maintenance goals
- 4x faster cycle pacing: 120s idle (was 600s), models train more often
- Enhanced reflection: quality model trains on ALL plan outcomes
- TOON encoder: 30-60% fewer tokens in prompts
- Benchmark solution accumulation: ground truth Rust code stored for Phase 3
Peer Discovery
- Blockchain peer discovery: ERC-8004 auto-mint, on-chain registry sync
- link_peer UPSERT: any node can register, not just Railway clones
- Startup discovery reads own children table + parent siblings
- Ghost cleanup no longer deletes reachable linked peers
Studio
- No polling: status fetched event-driven only
- System metrics: CPU/RAM/Disk in status bar
- Feedback buttons: text
good/bad, click-locks create_script_endpointremoved: cartridges only
v5.1.0 — Deep Planning + Cartridge Fix + Studio Polish
v5.0.0 — Three-Model Coding Intelligence
v4.0.0 — WASM Cartridge System
v3.4.0 — Major Structural Refactor
License
MIT