Skip to main content
RigorDesk logoRigorDesk
Start free

The chairman

The chairman is the model that reads the full debate transcript after the rounds end and writes the final verdict. It is separate from the panel. It never speaks during the debate itself, only synthesizes at the end.

Default chairman

The default chairman is deepseek/deepseek-v4-pro, chosen for:

  • Strong instruction-following (so the verdict format is consistent)
  • Low cost (the chairman runs once per debate, not per round)
  • Reasonable quality on synthesis tasks

You can override this per-debate. Set the chairman model when you start the debate via the assembly's formatMeta.chairmanModelId, or in the start request body.

The default works for 90% of questions. Override it when:

  • You need a stronger model for a high-stakes recommendation. Use Claude Opus 5 or GPT-5.6 Sol for the verdict.
  • You're running a cheap Refinement panel and want a frontier model to write the final answer. Use Opus 5 as chairman over a panel of Haiku + Mini.
  • You have a domain-specific need. For example, use a medical chairman model for clinical reviews when such a model becomes available.

What the chairman sees

The chairman's input is the full transcript:

  • The original prompt
  • The panel configuration (which models, which format)
  • Every round, in order, with the full text of every model's turn
  • The pairwise consensus scores for every round
  • The convergence status (did the panel converge? on which round?)
  • The disagreement arrays, what each model explicitly said it disagreed with

The platform explicitly tells the chairman:

  • Which format the debate used, so it knows whether to pick a winner like Oxford Union or write a unified verdict like Refinement
  • The user's original prompt, so it can answer the question instead of describing the debate
  • That it should produce a single best answer, not a meta-commentary about the debate
  • That it should preserve dissent, the strongest surviving counter-argument with attribution

What the chairman produces

The chairman's output is the finalAnswer field. Its structure varies by format:

  • Refinement / Devil's Advocate / Delphi / Steel-Manning. A unified answer to the user's prompt, optionally with per-claim confidence scores and a "remaining tensions" section
  • Oxford Union. A verdict that picks a side, proposition or opposition, explains why, and surfaces the losing side's strongest argument
  • Socratic. A revised version of the original position that incorporates concessions made during cross-examination
  • Red Team. A severity-graded remediation list, which is the judge's role

The chairman model also emits a confidence score for each major claim it makes in formats that support them. The verdict UI surfaces these scores.

How the chairman differs from the panel

Three things the chairman does that the panel models don't:

  1. Reads the full transcript. Panel models only see prior turns, not the consensus scores.
  2. Writes once. No rotation, no round-robin, no adversarial pressure. The chairman's job is synthesis, not debate.
  3. Has format-aware output. The chairman adapts its output to the format. A Refinement verdict reads like a recommendation. An Oxford Union verdict reads like a judge's ruling.

When the chairman fails

The chairman model itself can fail with a bad API response, a content filter, or a timeout. The platform retries the call. If the chairman call fails 3 times, the orchestrator marks the session FAILED and exposes the failure reason on the session row. The platform preserves the rounds' data. Only the finalAnswer stays missing.

This is the most common reason a debate completes with no finalAnswer despite running all rounds successfully. If you see this, the fix is to override the chairman to a different model and retry.

A note on bias

The chairman is a single model, so it has a single-model bias on the verdict. The platform mitigates this two ways:

  1. Dissent preserved. The platform explicitly instructs the chairman to surface the strongest counter-argument. This forces the chairman to acknowledge its own potential blind spots.
  2. The panel is the check. If the chairman is biased, the panel's per-model scores will reflect it. An outlier chairman verdict will carry low per-claim confidence from models that disagreed.

For most questions, these two checks are enough. For very high-stakes decisions, run the same debate twice with two different chairmen and compare the verdicts. If they agree, the conclusion is robust to chairman choice.