llm-provider-forensics

by andyrenxu7255

|

View Chinese version with editor review

安装

claude skill add --url https://github.com/openclaw/skills

文档

LLM Provider Forensics

Agent-facing forensic skill for identifying what an LLM endpoint most likely is.

Trigger conditions

Use this skill when asked to:

  • verify whether a claimed model is genuine
  • identify which family an endpoint most resembles
  • distinguish focused route vs wrapped route vs aggregation pool
  • compare multiple providers claiming to expose the same model
  • evaluate primary/fallback/avoid decisions
  • deeply audit suspicious gateways for GPT / Claude / Gemini / GLM / Qwen / Kimi / MiniMax / DeepSeek behavior

Core rule

Do not output false certainty. Produce a confidence-based operational judgment.

Coverage

Families:

  • OpenAI-compatible protocol layer
  • GPT / OpenAI-style
  • Claude / Anthropic-style
  • Gemini / Google-style
  • GLM / Zhipu-style
  • Qwen / Tongyi-style
  • Kimi / Moonshot-style
  • MiniMax-style
  • DeepSeek-style
  • mixed aggregation pool / compatibility gateway

Dimensions:

  • catalog topology
  • protocol compatibility
  • response schema shape
  • repeated stability
  • strict formatting control
  • family fingerprinting
  • long-context retention
  • structured-output stress
  • refusal/safety style
  • randomness / variance profile
  • streaming / error fingerprints
  • cross-protocol consistency

Current implementation note:

  • openai-compatible now means protocol layer only, not GPT-family proof.
  • The deepest automatic suite is strongest for OpenAI-compatible / mixed gateway providers.
  • Anthropic-native and Gemini-native routes currently have solid protocol/family checks, plus native deep tests, but protocol success alone must not be read as family proof.
  • Treat all family conclusions as confidence-based and inspect references before overclaiming.

Investigation workflow

  1. Identify likely protocol family or families.
  2. Probe catalog/list endpoints when available.
  3. Probe minimal inference endpoints for each plausible protocol family.
  4. Separate protocol-layer conclusion from suspected model family conclusion.
  5. Run repeated stability tests on the best working route.
  6. Run strict formatting tests.
  7. Run deeper advanced dimensions when the user prioritizes realism over speed.
  8. Inspect family fingerprint evidence and produce a confidence-based judgment.

References to load as needed

  • Main checklist: references/forensics-checklist.md
  • Advanced dimensions: references/advanced-dimensions.md
  • Error/stream/variance: references/error-stream-variance.md
  • Protocol specifics: references/protocol-openai.md, references/protocol-anthropic.md, references/protocol-gemini.md, references/protocol-glm.md
  • Family fingerprints: references/fingerprint-*.md
  • Native deep tests: references/deep-claude.md, references/deep-gemini.md

Final labels

  • high-confidence-focused-or-genuine-route
  • medium-confidence-likely-routed-or-wrapped
  • high-confidence-multi-model-aggregation-pool
  • low-confidence-or-unusable

Use high-confidence-focused-or-genuine-route sparingly. It should require:

  • stable repeated success
  • no strong mixed-pool signal
  • coherent family fingerprint
  • and no obvious gateway-normalization red flags in deep tests

Agent output contract

Return sections in this order:

  1. Declared facts
  2. Availability status
  3. Protocol-layer findings
  4. Suspected model-family findings
  5. Stability findings
  6. Capability/format findings
  7. Advanced-dimension findings
  8. Final judgment
  9. Need-human-review items
  10. Recommended operational posture

Preferred execution

bash
python3 scripts/llm_provider_forensics.py --config /root/.openclaw/openclaw.json --providers omgteam ypemc vpsai --model gpt-5.4 --deep