How it works

A session has four steps: discovery, handshake, inference and settlement. Nothing in the path is operated by GenKaKu Labs.

1. Discovery

A worker registers in WorkerRegistry with:

• a peer ID — the libp2p identity it will answer on,
• a policy hash — which model and rules the worker is offering,
• an optional locator — an address hint such as a multiaddr.

The registry exposes a public count() and isActive(address). The app reads them through a public RPC; there is no directory API to trust.

Locators are optional. The desktop app does not publish your public IP unless you explicitly turn publishAddress on, and a published locator is permanent once it is on-chain.

2. Handshake

The app opens a libp2p connection to the worker's peer ID. NAT traversal uses DCUtR and, when both sides are behind symmetric NAT, a relay for the handshake only. Relays do not carry model traffic: a 128 KiB / 2 min relay budget cannot sustain an LLM stream.

The hello exchange pins:

• the peer ID (must match the registry entry),
• the worker's wallet address,
• the policy hash,
• the hourly quote and billing mode.

3. Inference

The worker runs the model locally with Ollama and streams tokens back over the direct connection. Prompts are not stored by the marketplace; the worker operator can see the traffic it serves.

4. Settlement

Billing uses an hourly quote with started-minute accrual:

• the worker advertises a price per hour,
• each started minute counts as one minute,
• a two-minute job pays two minutes, not one hour.

The user opens a session in JobEscrow with an amount of test USDC. When the session ends:

1. the worker signs a receipt over the session ID, spend and a receipt hash;
2. settle() verifies that signature and releases the spend;
3. 99% goes to the worker, 1% to the treasury;
4. the remainder returns to the user with refund().

A worker can claim() its accrued balance at any time.

What replaces the orchestrator

Old model	GenKaKu
Server assigns jobs	User picks a worker from the contract
Server stores prompts	Direct peer connection
Server meters usage	Started-minute accrual, signed receipt
Server pays out	Escrow released on-chain

The trust assumptions are smaller because there is less in the middle. They are not zero: you still trust the worker with the prompt, and you still trust the code you run.