Define boundary, exclusions and source of truth.
Gateway and provider availability, maintenance, customer-caused events and external dependencies need explicit treatment.
Measure the service
before guaranteeing it.
No uptime, latency, throughput, support-response or credit commitment is active on the public planning site. A production SLA must be tied to a verified supplier route and measurement method.
Use percentiles and a reproducible workload.
Latency, time-to-first-token, throughput and error-rate targets must name model, route, traffic profile and measurement window.
State committed and burst limits.
Tokens per minute, requests per minute, concurrency, queue behavior and overage response must be written.
Severity, response, escalation and credits.
Contacts, support hours, incident notices, root-cause reporting and exclusive remedies belong in the signed service schedule.