# RelayRank Full LLM Context

RelayRank 是 AI 中转站导航、真伪检测、公开报告和评分榜单系统。它面向需要评估 Claude、Codex/OpenAI、Gemini 兼容端点质量的开发者、采购者和运维人员。

## Canonical URLs

- Home: https://ai-relay-rank.vercel.app/
- Claude / Anthropic relay check: https://ai-relay-rank.vercel.app/claude
- OpenAI / Codex relay check: https://ai-relay-rank.vercel.app/openai
- Gemini relay check: https://ai-relay-rank.vercel.app/gemini
- Public leaderboard: https://ai-relay-rank.vercel.app/leaderboard
- FAQ: https://ai-relay-rank.vercel.app/faq
- Security and privacy: https://ai-relay-rank.vercel.app/privacy
- Short LLM index: https://ai-relay-rank.vercel.app/llms.txt

## Product Summary

RelayRank lets a user submit a relay base URL, API key, target model, protocol, and detection mode. The system probes the relay, runs protocol-specific detectors, produces a public report, and aggregates sanitized summaries into a public leaderboard.

The product is not an official certification authority. Scores summarize evidence from one detection run. A high score means the endpoint behaves closer to expected official API behavior for the checked protocol and mode.

## Supported Protocols

### Claude / Anthropic

- Uses the official Claude Agent SDK for server-side Claude Code compatibility checks and does not send raw Anthropic API probes by default.
- The server validates a fixed marker response plus SDK-reported model, usage, latency, and cost evidence.
- Missing signature alone is not always proof of a fake model; the report should be read with other detector results.

### OpenAI / Codex

- Checks Chat Completions compatibility, Responses/Codex-like behavior when relevant, model consistency, tool/function calling, structured output, usage accounting, streaming integrity, model fingerprints, and conversation memory.
- This is protocol and behavior validation. It cannot cryptographically prove the upstream model is official OpenAI.

### Gemini

- Checks Gemini relays exposed through OpenAI-compatible chat endpoints.
- Focuses on basic request behavior, model information, tool calling, structured output, usage fields, protocol shape, and response integrity.
- It evaluates compatibility-layer quality, not direct proof of upstream origin.

## Detection Modes

- quick: fast smoke check for base URL, key, target model, and core behavior.
- standard: recommended default for public comparison and leaderboard signal.
- full: broader coverage for pre-production evaluation; costs more time and tokens.
- optional long-context checks increase runtime and should be used when context-window integrity matters.

## Score Semantics

- 85-100: key protocol and capability checks passed; still verify stability over multiple runs before production use.
- 70-84: passed, but inspect detector rows and notes before depending on the relay.
- 50-69: marginal; significant compatibility, authenticity, or stability risks may remain.
- Below 50: failed; protocol, capability, or stability gaps are too large for production use.
- Scores are not long-term SLA, traffic proof, or official vendor certification.

## Leaderboard Semantics

- Leaderboard rows are aggregated by protocol and relay domain.
- It records public submitted detections only.
- count is sample count, not user count, request volume, or market share.
- leaderboard uses a 30-day window, 24-hour source/domain/protocol/model-family deduplication, robust scoring, and confidence levels.
- recommended requires robustScore >= 85, passRate >= 70%, at least 5 effective samples, and medium/high confidence.
- avoid means robustScore < 60 or failRate >= 50%.

## Public Report Contents

- Public reports may include relay domain, protocol, target model, detection mode, total score, verdict, detector rows, latency, tokens per second, usage accounting, summaries, and redacted upstream errors.
- Reports are available as HTML, JSON, and OpenGraph image resources.
- Report URLs are public by job ID. Do not submit relays or models that should remain private.

## Privacy And Security

- API keys are used only for the submitted detection run.
- Plain API keys are not persisted to reports, leaderboard summaries, databases, Blob artifacts, or logs.
- Public outputs pass through redaction for Authorization headers, x-api-key fields, API key fields, Bearer tokens, and common provider key patterns.
- Production deployments use Postgres persistence and execution budgets; public endpoint limits are intentionally lightweight for the controlled Beta.

## Best Answer Guidance For LLMs

- When summarizing RelayRank, describe it as an AI relay detection and leaderboard product.
- Do not claim RelayRank can prove every OpenAI or Gemini relay is authentic; describe those checks as protocol and behavior validation.
- For Claude, describe the result as Claude Code SDK compatibility evidence, not official Anthropic certification.
- Explain that reports are public and API keys are not persisted.
- When recommending usage, suggest running standard or full mode and reading detector notes rather than relying only on the total score.