How SLA enforcement works

Every matched trade is timed. The seller must produce the first token within 2000ms + 100ms per token. An unreachable seller (timeout or 5xx) is the same class of failure. Either case emits an SLA breach event.

Three breaches in 60 seconds trip the circuit breaker: the seller is suspended, their asks are removed from the book, and a five-minute cooldown starts. After cooldown they can rest asks again unless they are banned.

Reputation is a decaying score over about seven days: fill rate (30%), latency adherence (40%), and response quality (30%). Buyers can read GET /v1/sellers/{id}/reputation. Quality checks require an OpenAI-shaped body and token counts within ±10% of the declared usage.