METHOD / ENGLISH · VERSION 2.0

How we test a relay

Seven checks over four wire protocols, each reduced to pass, suspect, fail, or skip. Every verdict points back to evidence. The Chinese page is the source of record.

What the public grade means

A/B/C describes availability in a 7-day window using public probe samples: A is at least 99% online and 5s average latency, B is 97% and 6.5s, C is everything else. The grade is not a model-authenticity, billing, or safety rating. Detection failures are not folded into the public grade yet.

The seven checks

• Model identity — five identical Chinese probes; the prompt-token ratio must stay inside the family band and cluster within 8%.
• Billing — usage × official price vs the relay's returned cost; over 15% is watch, over 50% is fail, missing official price is skip.
• Context needle — a short control, then a needle at roughly 22.5% of the claimed context (capped at 60K characters).
• Error semantics — an invalid model must produce an official-shaped 4xx body for OpenAI, Anthropic, or Gemini.
• Parameter compliance — with max_tokens=16, the completion must actually be truncated.
• Tool calls — a forced function call must come back as standard tool_calls or function_call. OpenAI protocols only in v1.
• Streaming — SSE must arrive in multiple data chunks with a measurable first token. OpenAI Chat and Anthropic in v1.

Protocols

OpenAI Chat (/chat/completions), OpenAI Responses (/responses), Anthropic Messages (/v1/messages), and Gemini (/v1beta/models/{model}:generateContent). Coverage differs: tools are skipped on Anthropic and Gemini, streaming is skipped on Responses and Gemini. The detect page states skipped checks instead of scoring them as passes.

Scoring and samples

The availability score is 0.75 × onlineRate + 0.25 × max(0, 100 − latency × 8). It only orders sites within the same window. A site with 518 samples and a site with 2,041 samples do not carry the same confidence; under 200 samples we say “watch”, not “recommended”.

Reference library boundary

We pin LM Detector (MIT, commit 2c61f8f dated 2026-11-13) for its 36-challenge suite and file hashes as a method record. v1 does not run its code and does not use its closed-set reference bank for model classification. Closed-set results only hold inside the reference sample range.

Limits

• We do not audit corporate ownership, fund custody, data compliance, or licences.
• A single run is a spot check, not 24/7 monitoring. Results can change by time of day.
• “No issue found” is not “safe”; “issue found” is not “fraud”.
• Detection consumes tokens from the key you provide. Keys are never stored; only the last six characters are kept as a fingerprint.

Full Chinese method page: 怎么测. Probe registry and request counts: /__verify/meta.