Scoring: price, latency, quality
The auction ranks effective cost, not sticker price. Latency and quality are measured by the router, not claimed by providers.
Price alone is the wrong thing to rank. A quote that saves you a cent and costs you two seconds of latency is not the best bid, and a provider that answers fast with the wrong model is worse than either. The auction therefore ranks an effective cost built from three terms: price, latency, and quality.
The three terms
| term | source | what keeps it honest |
|---|---|---|
price | the provider's signed quote | second-score clearing makes truthful quoting dominant |
latency | router-measured TTFT from the metering proxy | measured on real traffic, never self-reported |
quality | the provider's reputation score per model class | fed by canaries, audits, and challenge outcomes |
The measured terms matter more than they look. A marketplace that ranks on claimed latency invites optimistic claims; one that measures it on the requests you actually paid for makes the score self-enforcing. Slow providers lose auctions automatically, with no SLA lawyering.
Priorities: you choose the weights
The default weighting is price-dominant. Priority presets reweight the score per request: cheapest pushes price to nearly everything, fastest trades money for measured TTFT, and the balanced default sits between. The auction machinery is identical in every case; only the weights move. Preset details for customers live in Routing controls.
- router/src/auction.ts the effective-cost score and presets
- router/src/proxy.ts TTFT measurement on live streams