<!-- ABOUTME: Multi-Agent Debate (MAD) protocol enforcing epistemic validity, Response Anonymization, and weighted consensus. --> <!-- ABOUTME: Eliminates sycophantic conformity (85.5% modal error adoption) and self-bias in automated code review. -->

Epistemic Multi-Agent Debate (MAD) Review

Rules

1. Absolute Response Anonymization: Strip all agent names, model identities, role attributions, and conversation metadata from debate transcripts before feeding arguments to reviewer agents.
2. Asymmetrical Expert Personas: Assign distinct, specialized reasoning paths (Security Auditor, DB Architect, Performance Engineer, Compliance Officer) rather than homogeneous agent teams.
3. Directed Acyclic Graph (DAG) Topologies: Route feedback through structured, unidirectional review stages rather than unconstrained, cyclic peer-to-peer discussions.
4. Weighted Post-Hoc Voting: Avoid expensive, fragile forced consensus loops; aggregate independent reasoning paths through mathematically weighted scoring based on domain expertise.
5. The Invariant-Breaker Verification: Require reviewer agents to propose concrete, failing exploit test cases rather than passive prose summaries.

Workflow

1. Persona Initialization: Instantiate 2–3 distinct specialist personas based on the intent's mandatory_reviewers (e.g. security-privacy-agent for trust boundaries, backend-engineer-agent for persistence).
2. Anonymized Argument Generation: Each persona independently analyzes the diff and generates findings, risks, and proposed exploit tests without viewing peer outputs.
3. Anonymized Cross-Evaluation: If critiques diverge, present the competing arguments to each agent with all identity headers masked (e.g. [Proposal A], [Proposal B]).
4. Weighted Decision Aggregation: Aggregate votes where the primary domain expert holds 60% voting weight and secondary reviewers hold 40%.
5. Actionable Output Synthesis: Produce a structured JSON/YAML finding artifact containing validated defects, required patches, and new Invariant-Breaker test code.

Verification

• Verify that all review output artifacts are formatted as structured JSON/YAML with zero identity attribution.
• Confirm that any identified security or correctness defect includes a reproducible unit or integration test case.