ridwannurudeen/hermes-council/multi-perspective-analysis

Multi-Perspective Analysis

Use the council_query tool to submit complex questions for adversarial deliberation by five personas, each representing a distinct intellectual tradition. Use council_decision when the user needs a concrete recommendation between options.

When to Use

• Complex questions with no clear right answer
• High-stakes decisions where multiple viewpoints matter
• Research evaluation to stress-test findings
• Strategic planning to surface blind spots
• Contentious topics where bias is likely

The Five Personas

How to Use

Basic Query

council_query(question="Should we migrate our monolith to microservices?")

With Context

council_query(
    question="Is proof-of-stake more secure than proof-of-work?",
    context="We're evaluating consensus mechanisms for a new L1 blockchain targeting institutional users."
)

Subset of Personas

council_query(
    question="Will quantum computing break RSA by 2030?",
    personas=["skeptic", "oracle", "arbiter"]
)

Cost Modes

council_query(question="Should we rewrite the storage layer?", mode="fast")
council_query(question="Should we rewrite the storage layer?", mode="standard")
council_query(question="Should we rewrite the storage layer?", mode="deep")

• fast: Skeptic + Arbiter for cheap pre-checks
• standard: Full council for normal decisions
• deep: Full council plus a second Arbiter pass for high-stakes decisions

Decision Review

council_decision(
    options=["Ship now", "Delay for audit", "Cut scope and ship"],
    decision_context="Hackathon demo is tomorrow; auth changes are unreviewed",
    criteria=["demo impact", "safety", "reversibility"]
)

Interpreting Results

The council returns a structured verdict:

• confidence_score (0-100): Arbiter's posterior confidence after weighing all arguments
• conflict_detected: True if personas strongly disagree (confidence spread > 30%)
• persona_responses: Each persona's analysis with their individual confidence
• arbiter_synthesis: The Arbiter's Bayesian synthesis with prior/posterior reasoning
• action_summary: Recommendation, top risks, missing evidence, and next actions
• dpo_pairs: Preference pairs for training (aligned vs overruled responses)
• sources: Model-cited URLs
• verified_sources: URLs actually retrieved by the council evidence layer

Confidence Guidelines

• 80-100: Strong consensus, high-quality evidence
• 60-79: Moderate confidence, some unresolved tensions
• 40-59: Mixed signals, significant uncertainty
• 0-39: Low confidence, major disagreements or insufficient evidence

Workflow: Research Then Evaluate

For best results, combine web research with council evaluation:

1. Use web_search to gather initial data
2. Use web_extract to read key sources
3. Synthesize findings into a draft analysis
4. Use council_evaluate to stress-test your analysis
5. Revise based on council feedback

Custom Personas

Add custom personas in ~/.hermes-council/config.yaml:

personas:
  security_analyst:
    tradition: "Adversarial security thinking"
    system_prompt: "You are a security analyst. Find every attack vector..."
    tags: ["security", "adversarial"]

Custom personas are merged with defaults. Use the same name to override a default.