How it works

A request becomes a matched job, real inference, two layers of verification, and an on-chain payment — in a couple of seconds.

1RequestOpenAI-style call2Matchbest node wins3Servestream tokens4Verifysample & score5Settle95% / 5% on-chain

  1. Request

    You call POST /v1/chat/completions (OpenAI-compatible) with a Bearer API key. Quota and prompt-abuse limits run before anything touches the chain.

  2. Match

    The gateway normalizes the request to a canonical job and the matching engine picks a node by price and reputation.

  3. Serve

    The node streams tokens back over a WebSocket; the gateway proxies them to you and counts independently — you're billed on min(node, gateway) tokens.

  4. Verify

    Layer-B checks structure (non-empty, length, loop detection, the node is pinned to what it sent). Layer-A re-runs ~5% of jobs on oracle inference and compares embedding similarity — anomalies are flagged for review, never auto-slashed.

  5. Settle

    With an open credit session the job settles off-chain against your signed cap and flushes in one batchSettle; otherwise it's a per-job escrow release. Either way: 95% to the node, 5% to the protocol treasury.

  6. Reputation & economics

    The node's 5-dimension score updates and snapshots on-chain daily. The treasury sweeps fees 60/20/20 (ops / stakers / burn) — the burn shrinks supply.

See it live in the dashboard → Quickstart