Probe a live model endpoint once and get a signed latency/availability/consistency verdict vs the provider's claimed SLA.