Multi-Agent Consensus

Three LLMs scan in parallel, submit scores on-chain, aggregate via intelligence-weighted averaging.

Multi-Agent Consensus

Three different AI models analyze every package in parallel. Each submits an independent risk score on-chain. Scores are aggregated using intelligence-weighted averaging.

Agent Lineup

Agent	OpenRouter	OpenAI fallback
agent-1	Claude Sonnet 4	GPT-4.1
agent-2	Gemini 2.5 Flash	GPT-4.1 Mini
agent-3	DeepSeek Chat	GPT-4.1 Nano

Model diversity reduces single-model blind spots. Override via AGENT1_MODEL, AGENT2_MODEL, AGENT3_MODEL.

What Each Agent Does

1. Receives package source, package.json, version history, and any known CVEs
2. Produces structured JSON: risk score (0-100), risk level, reasoning, vulnerability list, supply chain indicators
3. Submits score on-chain via OPMRegistry.submitScore()

Each agent scores a version only once.

Intelligence-Weighted Scoring

When ARTIFICIAL_ANALYSIS_API_KEY is set, scores are weighted by each model’s Intelligence Index and Coding Index from the Artificial Analysis API. Higher-capability models carry more weight in the final score.

Without the API key, agents are weighted equally (simple mean).

Provider Routing

• OPENROUTER_API_KEY set → routes through OpenRouter (multi-model)
• Only OPENAI_API_KEY → falls back to GPT-4.1 family
• Force a provider: LLM_PROVIDER=openrouter or LLM_PROVIDER=openai