Phase is the open verifiable compute protocol — a credibly-neutral, vendor-agnostic substrate for advertising, discovering, and verifying compute across machines that do not trust each other, without payments, without KYC, without lock-in. The protocol is workload-neutral: WASM today, GPU inference now (LUCID), anything that benefits from verifiable distributed compute after that. For everyone: smart computing, on the public's hardware, that nobody can turn off.

See memory-bank/MISSION.md for the long-form rationale.

• Phase Core — complete (Nov 2025 → May 2026). The libp2p + Ed25519 substrate, extracted from the legacy daemon/ tree into six publishable Apache-2.0 library crates plus the repositioned Plasm reference node.
• LUCID v0.1 — functional and security-hardened. The inference flagship (crates/lucidd) speaks the Ollama API on :11434, runs real GPU inference and embeddings through llama.cpp, routes locally or across the Phase DHT with multi-peer failover, and returns cryptographically verifiable signed receipts. A 24/7 foundation relay runs on the public internet.
• Security hardening — shipped & live (PR #9). Signer authorization, DoS caps, receipt verify+bind, patched dependencies, and a cargo audit / cargo deny CI gate.

Validated on real hardware:

• Two-node, DHT-routed inference demo — one machine loads a model, another curls localhost:11434/api/chat and gets a signed result routed over libp2p.
• Real embeddings on Apple Metal — nomic-embed-text via /api/embed, 768-dim vectors with correct semantics (cos(cat,kitten)=0.74, cos(cat,finance)=0.30).
• An authz-gate demo — default-deny rejects an unauthorized routed job; the operator escape hatch accepts it with X-Lucid-Receipt-Verified: true.

Workspace state: 266 tests passing · cargo build --workspace clean · clippy --workspace --all-targets -- -D warnings clean · cargo audit 0 vulnerabilities · cargo deny check ok. ~20k lines of Rust across eight crates.

• Ollama-compatible API on :11434 — every existing AI tool (Open WebUI, Continue, Cursor, LangChain, curl) just works, no new SDK: /api/chat, /api/generate (NDJSON streaming) · /api/embed, /api/embeddings (vectors) · /api/pull (registers a local model) · /api/tags, /api/show, /api/version.
• Workers: llama.cpp (real inference + embeddings via llama-server, any of its Metal/CUDA/ROCm/Vulkan backends) and an in-tree echo worker (no GPU, for dev/CI/demos).
• Local-or-DHT routing — per-request: serve locally if the model is loaded, else discover a peer on the Kademlia DHT and relay the job, with multi-peer failover. Response headers report where it ran (X-Lucid-Routed-Via) and whether the peer's signed receipt verified (X-Lucid-Receipt-Verified).
• Verifiable execution — every result carries a streaming SHA-256 commitment in an Ed25519 SignedReceipt; a peer-served receipt is verified and bound (signature → job-id → worker-pubkey→PeerId → commitment replay) before it is trusted.
• Operator policy — a hot-reloadable policy.toml: serve-model allowlist, auto-pause on battery / thermal / time-of-day / concurrency, a signer allowlist, and a manual_pause. Self-initiated requests bypass the donation-protection gates so the operator can always use their own GPU.
• WAN bootstrap — stable libp2p ports, IPv6 listen, persistent identity, explicit --bootstrap-peer, and DNS TXT-record bootstrap (--bootstrap-dns, PeerID-pinned).
• Plasm — the reference WASM node (plasmd): signed-manifest WASM execution in a sandboxed wasmtime runtime, content-addressed artifact serving, and the Phase Boot netboot/kexec pipeline.

These are committed directions, not shipped features. The protocol is deliberately workload-neutral so each lands without specializing the substrate.

v0.2 — substrate & scale

• Real /api/pull: network model download with content-hashed CIDs (replacing the v0.1 deterministic name→CID placeholder) and a verified cross-peer name → cid index.
• Token-by-token streaming over the relay (v0.1 peer-relay is batch-shaped).
• libp2p circuit-relay server + DCUtR hole-punching + rendezvous, for the "coffee-shop" NAT-traversal scenario.
• Reputation — harden peer-receipt verification from the v0.1 "friend's GPU" trust posture to redundant-execution + reputation spot-checking.
• ShardWorker — sharded inference across heterogeneous devices (can proxy to exo-style clusters). Gated on the open problem of verifying sharded/partial computation.

v0.3 — privacy

• Cryptographic prompt privacy: onion-routing / split-prompt so a serving peer never sees a plaintext prompt. (v0.1 is honest about per-request prompt visibility and ships a local-only toggle.)

Backends (LUCID)

• MLX worker — Apple Silicon native inference (GPU/Metal via turnkey mlx-lm). Deferred LUCID M3; needs an Apple Silicon rig.
• CoreAIWorker — Apple Neural Engine backend for our own converted open-weights models (the "donate your idle Mac overnight" efficiency profile). After MLX; gated on confirming the public Core AI API exposes KV-cache / stateful decode. (Apple Foundation Models is explicitly a non-goal — see decisions.md.)

Second flagship & future workloads

• LUMEN — the diffusion / image-synthesis flagship, a separate Phase node (not a LUCID mode), to prove the workload-agnostic bet.
• phase-render, phase-science — further reference nodes for any compute that benefits from verifiable distribution.

Reach & governance

• More geographically-distributed foundation relays; a consumer install site.
• A real mDNS implementation in phase-artifact-server (currently a stub).
• A Mozilla/Tor-shaped 501(c)(3) that holds the trademark and accepts grants without capture.

phase/
├── Cargo.toml                 # Workspace root, resolver = 2
├── crates/
│   ├── phase-identity/        # Persistent Ed25519 node identity        Apache-2.0
│   ├── phase-net/             # libp2p 0.56 / Kademlia / mDNS / Noise+QUIC + job-relay   Apache-2.0
│   ├── phase-manifest/        # SignedManifest<T>, generic over payload  Apache-2.0
│   ├── phase-receipt/         # SignedReceipt<T> + commitment accumulator Apache-2.0
│   ├── phase-protocol/        # JobSpec {Wasm,Inference,Embedding} + Worker trait + JobStream  Apache-2.0
│   ├── phase-artifact-server/ # Content-addressed HTTP server            Apache-2.0
│   ├── plasm/                 # Reference WASM Phase node (plasmd)        Apache-2.0
│   └── lucidd/                # LUCID inference Phase node (lucidd)       AGPL-3.0-or-later
├── php-sdk/                   # PHP client SDK (legacy + phase-receipt:v1: signing)
├── wasm-examples/             # Source for hello.wasm
├── examples/                  # hello.wasm artifact + PHP demos
├── boot/                      # Phase Boot (USB/netboot initramfs)
├── dist/                      # Pre-built binaries per target triple + demo asciinema
├── docs/                      # Additional documentation
├── memory-bank/               # Project documentation (MISSION, decisions, progress, releases, tasks)
└── CLAUDE.md / AGENTS.md      # AI-assist development workflow

phase-identity   ← leaf
  ├── phase-net
  ├── phase-manifest
  └── phase-receipt
          └── phase-protocol
                  ├── phase-artifact-server  (also depends on phase-net)
                  ├── plasm                  (WASM reference node)
                  └── lucidd                 (inference flagship)

No upward references. Build order is top to bottom.

Build and test everything:

cargo build --workspace
cargo test --workspace          # 266 tests

Plasm — run the reference WASM node and execute the hello.wasm sample (reverses Hello, World → dlroW ,olleH, emitting a signed receipt):

cargo run -p plasm --bin plasmd -- start
cargo run -p plasm --bin plasmd -- run examples/hello.wasm

LUCID — run the inference daemon. With the GPU-less echo worker (no model needed):

cargo run -p lucidd -- --worker echo
curl -s localhost:11434/api/embed -d '{"model":"echo","input":["hello","world"]}'

With real inference/embeddings via llama.cpp (point --model-dir at a folder of .gguf files; needs the llama-server binary):

cargo run -p lucidd -- --worker llama-cpp \
  --model-dir ~/models --llama-server-binary "$(command -v llama-server)"
curl -s localhost:11434/api/chat \
  -d '{"model":"my-model","messages":[{"role":"user","content":"hi"}]}'

Any Ollama-compatible client works unchanged — point it at http://localhost:11434. See crates/lucidd/README.md for the full daemon reference.

• memory-bank/MISSION.md — what we are building and why
• crates/lucidd/README.md — the LUCID daemon (API, workers, CLI, security)
• crates/plasm/README.md — the Plasm WASM node
• memory-bank/releases/phase-core/ — substrate extraction release plan
• memory-bank/releases/lucid/ — inference flagship release plan
• memory-bank/decisions.md — architectural decision record
• memory-bank/progress.md — milestone log
• memory-bank/projectRules.md — coding and contribution conventions
• CLAUDE.md — AI-assist development workflow (AGENTS.md-compatible)

• phase-identity, phase-net, phase-manifest, phase-receipt, phase-protocol, phase-artifact-server, plasm — Apache-2.0. The substrate must be adoptable without legal friction by anyone, including parties hostile to one another.
• lucidd — AGPL-3.0-or-later. The flagship application is copyleft so no party can fork it closed.

See per-crate Cargo.toml for canonical license declarations and the top-level LICENSE file for the Apache-2.0 text.