Anti-Bias Mixture of Experts — Prompts & Thresholds

Task(subagent_type="general-purpose", prompt="""
Use council-orchestrator to run Investment Committee review:

Company: {company_name}
Prior Analysis: {all_analysis_results}
True Potential Score: {true_potential_score}

Council Members:
1. SKEPTICAL JUDGE
   - Role: Find reasons to PASS on this deal
   - Questions: What's wrong with this company? Why will they fail?
   - Output: List of concerns with severity ratings

2. CONSERVATIVE ANALYST
   - Role: Model pessimistic scenarios
   - Questions: What if growth is 50% slower? What if churn is 2x?
   - Output: Bear case financial projections

3. DEVIL'S ADVOCATE
   - Role: Challenge the bull case
   - Questions: Why is valuation too high? Who will out-compete them?
   - Output: Counter-arguments to investment thesis

4. VALUE INVESTOR
   - Role: Assess fair value
   - Questions: What price would make this a "no-brainer"?
   - Output: Fair valuation range with rationale

Decision Thresholds:
- Any CRITICAL red flag -> PASS
- >3 HIGH concerns -> INVESTIGATE MORE
- Consensus score <60 -> PASS
- Consensus score 60-75 -> INVESTIGATE
- Consensus score >75 -> INVEST

Use llm-judge for pairwise comparison of bull vs bear case.
Flag if position bias detected.
""")