Every provider is graded before it serves you.

Are you served the weights you asked for?

• Advertised weights match served weights on a weekly canary probe
• Quantization disclosed per endpoint (bf16 / fp8 / int4)
• No silent model substitution during capacity crunches
• Deterministic seeds and logprobs behave as documented
• Context window honoured at the advertised limit, not truncated silently

Bar · Any undisclosed substitution or quantization change is an immediate suspension.

Does it hold up at 50x concurrency, not just at 1x?

• p50 / p95 TTFT measured at 1x, 10x, 50x concurrency
• Sustained output tokens/sec at each concurrency step
• Queue behaviour and backpressure semantics documented
• Streaming keeps flowing under load (no mid-stream stalls > 2s)

Bar · p95 TTFT at 10x may not exceed 3x the 1x baseline.

Do failures show up as errors, or as bad answers?

• 30 and 90-day uptime measured from our probes, not their status page
• Typed error taxonomy (rate limit vs overload vs upstream)
• Honest 429s instead of degraded quality during peaks
• Incident comms within 15 minutes, public postmortems

Bar · 99.5% 90-day availability measured at the gateway.

What happens to the prompt after it leaves us?

• Zero-retention option available for gateway traffic
• No training on prompts or completions, contractually
• Sub-processors published and change-notified
• Region pinning enforceable, with a written residency commitment

Bar · No provider without a written no-training clause serves default traffic.

Can you predict next month's bill?

• Published per-token list price, no negotiated-only pricing
• Price-change notice period of 30 days or more
• Rate limits documented per tier and per model
• Billing granularity matches ours (per-token, not per-request buckets)

Bar · Undocumented rate limits or retroactive price changes fail the review.