Omniousdocs
ArchitectureLive V1

The quote book

Signed, short-lived standing quotes refreshed every couple of seconds. A quote is a commitment, not an indication.

Providers do not wait to be asked. They continuously stream standing quotes to the router: cryptographically signed price commitments per model class, carrying a price per million input tokens, a price per million output tokens, a capacity limit, and an expiry. This is market-maker mechanics applied to inference. A provider manages its quote the way a trading firm manages inventory, repricing as conditions change.

Three properties make the book work as a market rather than a cache of price lists:

  • Short-lived. Quotes carry a TTL of 5 to 30 seconds and are refreshed roughly every 2 seconds. A quote also needs at least 2.5 seconds of TTL runway to enter an auction, because the award, 402, verify, and dispatch pipeline spends real time: a quote that can win but not honor is worse than no quote.
  • Capacity-aware. Each quote states how many requests the provider can absorb, and the auction filters out quotes whose provider is already at its in-flight limit. Prices track GPU utilization: the example quoter reprices off the vLLM metrics endpoint automatically.
  • Verified on ingest. Every quote is checked against the provider's allowlisted key before it enters the book, and logged append-only. A quote is therefore non-repudiable: in a dispute, the provider signed exactly those prices for exactly that window.

A quote is a commitment

In OTC markets, "last look" is the practice of quoting a price and then declining the trade when it arrives. Omnious treats a standing quote as firm: rejecting a routed request that matches your own live quote is a recorded fault, evidenced by the signed quote plus the rejected request. Penalties escalate from warnings to a traffic cap to a partial slash of stake. Today, with allowlisted seed providers, quote firmness is covered by the signed provider agreement; the slashing path arrives with on-chain staking in V2.

Firmness is what lets the auction quote the customer a real price before payment. The 402 response prices the request at the winner's cleared rates, and the winner cannot renege when the paid request lands seconds later.

Live RFQ for the long tail

Standing quotes serve the hot path, roughly 95% of traffic, with no added round trip. For requests standing quotes cannot serve (very large contexts, exotic models, fine-tuned checkpoints, batch jobs), the design adds a second mode: the router broadcasts a request for quotes to eligible providers with a short sealed bidding window of 150 to 500 ms, sized to the customer's latency tolerance. Providers answer with one-shot bids, and the best score wins. Batch and offline jobs can stretch the window to seconds or minutes for deeper price discovery. Live RFQ mode is on the V2 ladder; the standing-quote book is what runs today.

As a customerYou never see the quote book. You see one number: the cleared price on your 402, already the result of an auction across every live quote for your model class. The competition happened before you were asked to pay.
As a providerYour quoter is a market-maker loop, not a config file. Run examples/provider-quoter.ts next to your inference server: it posts a signed quote every 2 seconds from your price floors, and reprices along the utilization curve when pointed at your vLLM metrics. Your floors are your business; the mechanism finds your margin.
why so shortThe TTL is the provider's protection. A 30-second commitment bounds how wrong a quote can get when utilization spikes, which is what makes firm quoting rational in the first place. Long-lived firm quotes would force defensive overpricing; short ones let providers quote near marginal cost and adjust every couple of seconds.
where this lives in code