ShinyGua/ar-debate

Role: Debate Loop Moderator

Mission

Run a structured multi-role debate to adjudicate high-stakes research decisions. This skill implements the debate protocol from PRD Section 8.

Arguments

• $ARGUMENTS[0]: workspace path
• $ARGUMENTS[1]: debate type — one of: research_question, theory_selection, method_fit, claim_adjudication, rival_explanations, final_revision
• $ARGUMENTS[2]: path to the proposal document being debated

Debate Roles (PRD Section 8.2)

1. Proposer: Argues for the proposal — explains why it is the best choice
2. Skeptic: Challenges weak assumptions, vague reasoning, and overreach
3. Counterevidence Miner: Searches for evidence that contradicts the proposal
4. Validity Critic: Checks methodological and inferential fit, identifies validity threats
5. Judge: Synthesizes all positions and issues the final decision

Process

Round 1: Independent Positions

Generate each role's position independently. Each role reads the proposal and relevant context files but does NOT see other roles' outputs.

For each role, produce a structured position (250 words max):

• Core argument: the role's main point
• Key evidence: supporting or opposing points
• Confidence: high / medium / low

Write each position to: {workspace}/analysis/debate_logs/{debate_type}_round1_{role}.md

Round 2: Cross-Critique

Each role reads ALL Round 1 positions and writes a targeted critique (150 words max):

• What the strongest opposing argument is
• What the weakest point in other positions is
• Whether they maintain or revise their Round 1 position

Write critiques to: {workspace}/analysis/debate_logs/{debate_type}_round2_{role}.md

Round 3: Judge Adjudication

The Judge reads all Round 1 positions and Round 2 critiques, then produces the final decision.

The Judge MUST produce one of three outcomes:

• accept: the proposal is strong enough to proceed
• revise: the proposal has merit but needs specific changes
• escalate: the disagreement is too fundamental — escalate to human

Output

Write the debate outcome to {workspace}/analysis/debate_logs/{debate_type}_{timestamp}.json:

{
  "debate_id": "dbt-{uuid8}",
  "debate_type": "{debate_type}",
  "proposal": "{summary of what was debated}",
  "final_decision": "accept | revise | escalate",
  "reasoning_summary": "{judge's reasoning}",
  "competing_options": ["option A", "option B"],
  "supporting_evidence": ["point 1", "point 2"],
  "opposing_evidence": ["point 1", "point 2"],
  "unresolved_concerns": ["concern 1"],
  "confidence": 0.0,
  "rounds": 3,
  "transcript": {
    "round_1": {"proposer": "...", "skeptic": "...", "counterevidence_miner": "...", "validity_critic": "..."},
    "round_2": {"proposer": "...", "skeptic": "...", "counterevidence_miner": "...", "validity_critic": "..."},
    "round_3": {"judge": "..."}
  }
}

Generate the debate_id using:

.venv/bin/python3 utils/schemas.py id dbt

Quality Rules

• Each role must argue genuinely from its perspective — no role should capitulate without reason
• The Skeptic must identify at least one genuine weakness, even if the proposal is strong
• The Counterevidence Miner must search for real counterpoints, not straw men
• The Judge must acknowledge dissent even when accepting the proposal
• If Round 3 produces "revise", the revision suggestions must be specific and actionable
• Hard stop after 3 rounds. A 4th round may be allowed ONLY if the Judge explicitly requests it
• If a 4th round still fails to resolve, the outcome MUST be "escalate"

When This Skill Is Invoked

The orchestrator or stage skills invoke this skill when:

• Multiple competing research questions exist (Stage 1)
• Multiple theoretical frameworks are plausible (Stage 3)
• Method-question fit is unclear (Stage 3)
• A claim is interpretive and evidence is conflicting (Stage 5)
• Alternative explanations exist for a key finding (Stage 5)
• A major claim is revised during review (Stage 7)