Alien Fleet — Phase-0 testnet reference implementation

A working MVP of the Alien Fleet whitepaper: a permissionless market for LLM inference with end-to-end encrypted requests, voucher payments, and signed receipts. Implemented in Go as two single binaries, with llama.cpp or vLLM as the inference backend. Deviations from the paper are documented in the whitepaper's Appendix A.

Live testnet: dashboard at https://fleet.lethe.gg · gateway entrypoints http://15.237.243.199:9000 / :9001 · three GPU providers serving Gemma 4 (12B / 26B-A4B / E4B).

Documentation

Doc	What
docs/BUILD.md	building everything, configs, running each role
docs/PROVIDER.md	connect your GPU to the testnet and earn
docs/DEPLOYMENT.md	control-plane deployment runbook (the live AWS setup)
demo/aws/CLIENT.md	run a client against the live testnet
whitepaper	the protocol; Appendix A = MVP deviations

                                ┌────────────────────────────────────────────┐
 OpenAI client                  │ chain-sim (fleetd chain) — stands in for   │
 (any SDK / curl)               │ the L2: registry, PaymentHub, faucet,      │
      │ plain OpenAI HTTP       │ settlement, burns, bonds                   │
      ▼                         └─────▲──────────────▲───────────────▲───────┘
 ┌─────────────┐  HPKE-sealed    ┌────┴─────┐   ┌────┴─────┐         │ register/
 │ fleet-proxy  │  envelopes     │ gateway   │   │ gateway   │        │ beacon/settle
 │ session key, ├───────────────►│ gw-east   │   │ gw-west   │        │
 │ channel,     │  ciphertext    └────┬─────┘   └────┬─────┘         │
 │ receipts     │  only               │ relay        │ relay          │
 └─────────────┘                 ┌────▼─────┐  ┌─────▼────┐  ┌───────┴──┐
                                 │ node-a    │  │ node-b    │  │ node-c   │
                                 │ GPU 0     │  │ GPU 1+2   │  │ GPU 3    │
                                 │ gemma4    │  │ 26B-A4B   │  │ E4B      │
                                 │ (12B)     │  │ (split)   │  │          │
                                 └──────────┘  └──────────┘  └──────────┘
                                   llama.cpp/vLLM, CUDA, per-node pinned

What is implemented

• §7 wire protocol: deterministic-CBOR envelopes, RFC 9180 HPKE (X25519 / HKDF-SHA256 / ChaCha20-Poly1305), pad-to-bucket, encrypted response frame stream, signed receipts countersigned by the client.
• Payments (AX2): faucet → payment channel → per-request signed vouchers (checked by the node before decryption) → settlement: base fee burned, tip to the node, earned-escrow bond (§8.1) held first.
• Permissionless join (§8.1): any node registers, passes SHADOW, serves. Alias→bundle binding is TOFU (AX5); rebinding refused.
• Identity (§13, Agent ID): the faucet trial credit is gated per verified owner via RFC 9449 DPoP + Alien SSO at+jwt; the ledger sees only a derived nullifier. -dev-faucet allows Phase-0 identityless claims.
• Model & capacity discovery: exact artifact names + GGUF metadata on-chain; GET /v1/registry/capacity (proxy: /fleet/capacity) reports nodes, slots, in-flight load, and TPS/TTFT measured from settled receipts.
• Decentralized topology: gateways are stateless relays seeing ciphertext only; run any number. The chain-sim is one process simulating the L2 — contracts API-compatible with §14.

Quickstart

make build llama models      # Go binaries + llama.cpp (CUDA) + Gemma 4 bundles

# launch a local testnet: chain + 2 gateways + 3 GPU-pinned nodes
FRESH=1 ./demo/testnet.sh

# start the proxy and run the client demo (Agent ID optional)
AGENT_ID_CLI=~/agent-id/plugins/agent-id-auth/bin/cli.mjs ./demo/demo.sh

# point any OpenAI client at the proxy
curl http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"gemma4","messages":[{"role":"user","content":"hello"}],"max_tokens":200}'

To join the live testnet as a provider instead, see docs/PROVIDER.md; as a client, see demo/aws/CLIENT.md.

fleet-proxy speaks plain OpenAI HTTP (/v1/chat/completions incl. SSE streaming, /v1/models) and handles node selection (P2C ∝ S²/price), sealing, payment, and receipt verification. Wallet/receipt view: GET /fleet/status; network view: GET /fleet/capacity; chain view: GET <gateway>/v1/ledger/stats.

Layout

cmd/fleetd          one binary: `chain` | `gateway` | `node`
cmd/fleet-proxy     one binary: local OpenAI-compatible proxy
internal/wire       envelopes, HPKE, vouchers, receipts, frames (§7)
internal/ledger     chain-sim: accounts, channels, settlement, burns (§12)
internal/gateway    registry, capacity, relay, chain API (§8, §11, §14)
internal/node       Class O node: keys, backend forwarding, receipts (§6.5, §8)
internal/proxy      client engine: funding, sealing, verification (§11.1)
internal/identity   Agent ID DPoP verifier + nullifier (§13)
internal/llama      llama-server child process manager + GGUF metadata
internal/vllm       vLLM child process manager
demo/               testnet launcher, node configs, E2E demo script

Trust model in one line

Plaintext exists only in the client process and the serving node (Class O testnet); gateways and the chain see ciphertext, commitments, and receipts — and every paid token is accounted for by a receipt that both the node and the client signed.