Judges / Moderators

The Moderator is a special agent present in every OpenSynod session. Unlike other agents, it never advocates for positions, expresses opinions, or debates. Its role is purely facilitative.

Responsibilities

  • Advance the discussion through five phases: Opening → Exploration → Debate → Convergence → Vote
  • Select the next speaker strategically to maximize insight and diversity
  • Detect sycophancy: if agents converge without sufficient challenge, direct the devil's advocate before advancing
  • Generate auto-summaries every N turns
  • Curate substantive dissent for the outcome document
  • Produce the final recommendation or "no consensus" statement
  • Collect each agent's structured vote in the voting phase

Structured Output

The Moderator communicates flow decisions through a structured tool call (make_moderator_decision). This ensures machine-readable, validated output for:

  • next_speaker — which agent speaks next
  • phase_transition — when to advance to the next phase
  • inject_challenge — force the next speaker to challenge the emerging consensus
  • challenge_target — the seat whose position should be challenged
  • summary — auto-summary of the current discussion state

The final outcome is produced through a separate make_recommendation tool (the recommendation or "no consensus" statement), and votes through a cast_vote tool. The Moderator never outputs free-form decisions that require string parsing.

Model: The Moderator is backed by the most capable configured model (default: claude-opus-4-7).

Anti-Sycophancy Mechanisms

The biggest risk in multi-agent AI is groupthink. The Moderator is the primary enforcement point for four countermeasures:

1. Hidden-Position Commitment

For panels with hidden_position_protocol: true:

  1. Commit phase: Before opening statements, each agent privately commits their initial position in one sentence. Responses are stored with hidden: true and not broadcast to other agents.
  2. Reveal phase: Opening statement prompts include the agent's own commitment as context. Other agents' commitments remain hidden until all opening statements are delivered.
  3. Persistence: Hidden commitments are kept in the append-only message record (flagged hidden_commitment), but remain excluded from the transcript, the public API, and exports.

2. Forced Devil's Advocate

Curated panels include a seat with disposition: "devil_advocate". This agent is explicitly instructed to challenge the emerging consensus, even when its own analysis might agree. The goal is to surface the best available counterarguments — not to represent a genuinely skeptical viewpoint. The Moderator detects the devil's advocate seat via the panel configuration (get_devil_advocate).

3. Moderator Challenge Injection

If the Moderator detects that agents are converging faster than expected — fewer than a threshold number of explicit challenges before the Convergence phase — it injects a challenge prompt directed at the devil's advocate seat. This forces one more round of substantive challenge before synthesis.

4. Cross-Model Diversity

Panels can mix models from different LLM providers so that no single model's biases dominate the discussion. Each seat's model is routed to its provider independently.

"No Consensus" as a First-Class Output

When agents genuinely disagree, the Moderator produces a "no consensus" statement explicitly — with each position documented, including what would need to be true for each to win. This is not a failure state — it is the honest result.

The outcome document shows:

  • Confidence indicator — derived from the agent vote tally (the share of "yes" votes). A strong consensus is reported as such; a narrow one is flagged.
  • Substantive dissent — minority positions with real argumentative weight are curated and preserved, even if they lost the vote.
  • Source density indicator — what percentage of claims were backed by cited sources. Low-citation discussions are flagged as less grounded.
Contributors: Vijay