Operating facts for the gateway.
Monitoring scope, serving regions, data retention, access controls, and support channels. The service currently has no uptime SLA.
What the public status measures
A Cloudflare Worker checks public readiness and the model catalog every five minutes. Status reports the latest probe separately from rolling probe success and publishes the same data as JSON.
This is a readiness check, not an authenticated completion benchmark or an uptime service-level objective.
Processing and retention
- Serving regions
- the EU (Qwen3.8 27B) and the United States (Gemma 4 31B). The requested model determines the processing region.
- Purchased-credit prompts
- Processed in memory and not written to durable storage by the serving stack when the training setting is off.
- Operational logs
- Status, latency, error class, and calling IP; no request content. Retained up to 30 days.
- Billing metadata
- Model, token counts, timestamps, key prefix, and cost; no request content. Retained up to 12 months.
Account and control boundaries
- Inference and model discovery require a revocable bearer key.
- Customer keys cannot call billing, key-management, or administrative endpoints.
- Administrative surfaces use a separate origin and service-token access control.
- Prompt-cache entries are isolated by account.
- Billing records carry token counts and cost, not prompt or completion text.
Incidents and production qualification
Support is handled by email without a published response-time SLA. Production workloads should be qualified before launch with model, expected concurrency, and token volume.
Contact support@tiyuvta.ai. Benchmark dates and methods are published on the models page; contract terms are on the terms page.