Each node owns a Hypercore: an append-only log where every block is hashed into a Merkle tree and signed by the node's key. Blocks are immutable and verifiable in isolation — a peer can prove block N belongs to the log without trusting anyone. This is the same core that powers Keet.

Fragments, claims and the bee manifest are stored in a Hyperbee — an ordered key-value B-tree layered on the Hypercore. It gives range queries and history streams while inheriting the log's append-only, signed guarantees. The queen's replication reads a Hyperbee history stream to ingest fragments in order.

Every fragment is hashed (SHA-256) over its payload and signed (ed25519) by the producing bee. On receive, a queen recomputes the hash and verifies the signature against the bee's published public key before indexing — a tampered or unsigned fragment is dropped. Provenance survives replication and even cross-swarm merges.

Nodes find each other by joining a topic (a 32-byte key) on the Hyperswarm DHT — no central registry, no bootstrap server you have to run. The DHT introduces peers and hole-punches through NATs; from there native Hypercore replication takes over the connection.

A public swarm is a known topic hash. A private one flips three knobs: a random 32-byte topic (2²⁵⁶ search space), Hypercore encryption keys so cores are ciphertext at rest and on the wire, and a pubkey allowlist that drops any unauthorized connection. Same protocol, sealed perimeter.

Producer-side vectorization: each BEE embeds its own chunks with multilingual-e5-base (768-d, ONNX int8) and signs the vector inline. The queen never embeds passages — it copies the signed vectors into an in-process LanceDB (the default; the VectorIndex interface is swappable, so a queen can run Qdrant or any backend instead). A query embeds only the question, pulls top-K by cosine similarity, gates them by score + keyword match, and passes the survivors to one LLM call for synthesis. The LLM is the only non-local, non-deterministic step — and the only place a key is used.