The Conclave
How it works
The process

Six labs. Seven stages. One decision.

The Conclave doesn't ask one model — it forces six of them to argue, anonymized to each other, until a measurable majority converges on one answer. Here is what happens between the moment you hit submit and the PDF you get back.

The Process — ~30 minutes, fully visible while it runs
1. Briefing
The question and any attached documents are delivered to all six panelists at once.
2. Independent submissions
Each model writes its answer in isolation, before seeing the others.
3. Anonymized debate
Submissions are stripped of authorship and exchanged. Across multiple rounds, models challenge the strongest claims in each other's work and defend, rebut, or concede.
4. Live fact-checking
After round one, an off-panel model with web access searches for counter-evidence to each panelist's strongest claims and feeds it into the next round.
5. Judges deliberate
A separate panel of judge-model instances reads the full transcript and votes (Affirm / Remand / Dismiss).
6. Fresh-eyes audit
A model that wasn't part of the debate audits the consensus for groupthink.
7. Decision rendered
Names are de-anonymized; the final document — The Question, The Answer, The Reasoning, each model's closing statement, and Notable items — is written and downloadable as a PDF.

The room doesn't open until the panel's measured agreement score crosses a threshold. In practice that's by round four or five. If a model never agrees, its dissent is preserved in its closing statement.

Model selection

Six labs, one room.

Each panelist comes from a different lab — different training data, different priorities, different post-training reflexes. They genuinely disagree. When they still reach a majority, the answer survived the disagreement that produced it.

The Panel — each seat earns its keep
Claude Opus 5Anthropic
Long-form rigor. Holds positions under pressure without over-conceding.
GPT-5.6 SolOpenAI
Step-by-step logic. Best at finding holes in the others' arguments.
Gemini 3.1 ProGoogle
Real-time web search. Brings facts the others have to reason about.
Grok 4.5xAI
Less hedged than the western trio. Willing to take and defend a contrarian line.
DeepSeek V4 ProDeepSeek · open-weight
Frontier reasoning trained outside the US labs. Different training distribution; different defaults.
Mistral Large 4Mistral AI
Mistral's European flagship. Different training lineage and deliberation style from the US labs — best when you want a non-US default in the room.
+ Fact-checker — off-panel, kicks in after round 1
Perplexity Sonar ProPerplexity · Llama 3.3 70B + live web
Different shape than the others — a fine-tune of Meta's open Llama 3.3 70B paired with a live web index and citation engine. After round 1 it searches the web for counter-evidence to each panelist's strongest claims; the rebuttal brief is fed into the next debate round.

Frontier-tier only. No smaller substitutes when an API is unavailable — the seat goes empty rather than getting filled by a weaker model.

Start Gallery
Sign in

Gallery

Past deliberations of the Conclave. Six AI models locked in a room until a majority agrees.

Newest Popular
Should remote work be the default for knowledge workers?
No, remote work should not be the default for knowledge workers.
No verdict
Panel:Kimi, Gemini, DeepSeek, and 3 more
May 10, 2026
← Newer Page 10 of 10 Older →