Debate Protocol for MoE Classification System.
Implements multi-round debate when judges disagree, based on MAJ-EVAL in-group debate protocol (Chen et al., 2025) and CODITECT research.
Key concepts:
- MAX_DEBATE_ROUNDS: Maximum rounds before forcing consensus (default: 3)
- CONVERGENCE_THRESHOLD: Agreement ratio to stop debate (default: 0.8)
- Debate context includes other judges' positions and evidence
File: debate.py
Classes
Verdict
Judge verdict options.
JudgeEvaluation
Evaluation from a single judge.
Disagreement
Represents a specific area of disagreement between judges.
DebateRound
Record of a single debate round.
DebateConfig
Configuration for debate orchestration.
DebateResult
Result of the complete debate process.
DebateOrchestrator
Orchestrates multi-round debate when judges disagree.
Functions
create_default_orchestrator()
Create a debate orchestrator with default configuration.
requires_debate(evaluations, threshold)
Quick check to determine if debate is needed.
set_evaluation_callback(callback)
Set callback for conducting evaluations during debate.
get_debate_summary(result)
Generate a summary of the debate process for audit trail.