A structured deliberation engine
One question.
No easy answers.
Instead of asking a single AI model and trusting its answer, Conclave convenes a panel of independent models, forces them to argue, and subjects their conclusions to judicial review. Consensus must be earned through evidence — not assumed through agreement.
The eight phases
From question to verdict
Every Conclave runs the same rigorous sequence. Each phase surfaces blind spots, stress-tests positions, and drives toward a conclusion that has survived real opposition.
Briefing
The room is convened
Your question is parsed and the panel is assembled. Six advocates are selected — each a different AI model. They are seated, anonymized, and handed the same briefing. No model knows who else is on the panel.
Independent Submissions
Six sealed answers — written in isolation
Every advocate drafts their answer from scratch, without seeing anyone else’s work. The goal is to surface genuine differences in reasoning — not polished consensus. Each submission includes a falsifiable hypothesis, evidence labeled by type, and an honest self-assessment.
If one model fails, its seat is lost. The session continues as long as at least two submissions survive.
Collation
Authorship stripped. Ideas remain.
The orchestrator collects all submissions and removes every identity marker. From this point forward, every advocate reads arguments attributed to Advocate-A through Advocate-F — not to specific AI brands. This forces models to evaluate ideas on their merits, not their source.
Cross-Examination
Every position is challenged directly
Each advocate reads all anonymized submissions and writes a pointed challenge — not a polite review, but targeted contestation of specific claims by alias. All six fire in parallel. The strongest submissions get the hardest questions.
At the same time, Conclave brings in real-world counter-evidence the panel may have missed. This is the only point where live external evidence enters the deliberation.
Debate
Seven rounds of adversarial argument
Advocates argue across seven rounds. In each round, every advocate must make a declared move: Defend the original position with sharper reasoning, Concede a specific point to a challenger, or Revise the stance — citing explicitly what changed.
Between rounds, the system runs a stability analysis on each advocate’s position to detect sycophantic drift — models changing their answer to keep the peace rather than because new evidence compelled it. Midway through, the bench also performs a judicial checkpoint before the debate continues.
Judgment
An independent bench reads the full record
Judges — models that never participated as advocates — receive the entire transcript: submissions, challenges, every debate round, and the stability report. They independently evaluate whether convergence happened because of evidence, or just social pressure.
Each judge produces a structured opinion: a fact-check table with evidence quality ratings, a ranking of advocate positions, and a verdict — ACCEPT, SYNTHESIZE, or REMAND. A majority opinion is then written, and dissenting advocates may file a rebuttal.
Fresh Eyes
An outsider audits the consensus
An uninvolved model — one that has seen nothing of the deliberation — reads the majority opinion cold and audits it for groupthink, missing perspectives, and internal contradictions. This is a sanity check, not a veto.
Decision
The verdict is rendered
The deliberation is written up as a readable narrative, a structured verdict, and a smoke color representing the panel’s agreement level. A PDF is generated with the full record, evidence matrix, and majority opinion. The result is then delivered to you.
Reading the outcome
The smoke signal
Conclave’s verdict comes with a visual indicator of how strongly the panel agreed — so you know immediately whether the conclusion is solid ground or contested territory.
White Smoke
The panel reached a clear, strong consensus. The strongest position won cleanly and dissent was minimal or absent.
Gray Smoke
A majority agreed, but meaningful dissent exists. The verdict is defensible — but the minority view is worth reading too.
Black Smoke
The panel was split. No single position commanded enough support. The question may need reframing, or genuinely resists resolution.
Design principles
Why it’s built this way
Every structural choice in Conclave is meant to counter a specific failure mode of AI reasoning.
Isolation before debate
Advocates write their first answer before seeing anyone else’s. Isolation forces independent commitment and prevents early anchoring.
Anonymized authorship
Once submissions are collated, arguments are judged on logic and evidence — not on the reputation of the model that made them.
Sycophancy detection
The system tracks position changes across all rounds. If a model shifts without citing new evidence, that drift is flagged rather than rewarded.
Real-world evidence injection
Conclave deliberately introduces external counter-evidence at cross-examination, grounding the debate in the world rather than in isolated model reasoning.
Separate judges from advocates
The bench is made up entirely of models that never argued a case. Judges evaluate the record without having to defend their own earlier position.
Declared moves only
In debate, every advocate must state whether they are defending, conceding, or revising. Vague hedging is not a valid position.