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 4.8Anthropic
Long-form rigor. Holds positions under pressure without over-conceding.
GPT-5OpenAI
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 4xAI
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.
Kimi K2.6Moonshot AI · open-weight
Trillion-parameter mixture-of-experts with extended reasoning. Long-context champion — best when the question has many moving parts. Different RLHF lineage than the western models.
+ 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
What are the primary reasons people hire Harvey.ai? Discuss the key benefits and common use cases.
Organizations hire Harvey.ai for a bundle of interdependent reasons: initial procurement is driven by innovation signaling and FOMO, the investment is justified using marketed promises of attorney-hour recovery, and sustained ad…
Gray smoke
Panel:Claude, DeepSeek, GPT-5, and 3 more
May 13, 2026
Will Docusign experience a significant decrease in its customer base by 2028? Include a probability estimate, key factors influencing customer retention, and what developments could alter this outlook.
No, Docusign will not experience a significant (>10%) cumulative decrease in its customer base by 2028.
Gray smoke
Panel:DeepSeek, GPT-5, Gemini, and 3 more
May 13, 2026
What are the primary reasons people hire Harvey.ai? Discuss the key benefits and common use cases.
Organizations hire Harvey.ai primarily to achieve substantial, self-reported time savings on routine, document-heavy legal workflows—especially drafting, due diligence, and contract corpus Q&A—based on the expectation of efficiency gains, n…
Gray smoke
Panel:Magistral, Claude, GPT-5, and 3 more
May 13, 2026
What are the main reasons people hire Harvey.ai? Include key benefits and any common use cases.
People hire Harvey.ai primarily to recover measurable attorney time—2 to 10 hours per week per lawyer—through two core use cases: first-draft generation of legal documents and fast contract/knowledge querying.
Gray smoke
Panel:Claude, Grok, GPT-5, and 3 more
May 13, 2026
By 31 Dec 2028, will Amazon’s custom AI accelerators (Trainium + Inferentia) account for at least 10% of global data center AI accelerator revenue (training + inference), excluding general-purpose CPUs? Include a probability estimate, key uncertainties, and factors that could influence this outcome.
Amazon’s Trainium and Inferentia accelerators will likely account for 8–12% of global data center AI accelerator revenue by 31 December 2028, with a 55% probability of reaching or exceeding 10%.
Gray smoke
Panel:GPT-5, DeepSeek, Magistral, and 3 more
May 13, 2026
By 31 Dec 2028, will Amazon’s custom AI accelerators (Trainium + Inferentia) account for at least 10% of global data center AI accelerator revenue (training + inference), excluding general-purpose CPUs? Include a probability estimate, key uncertainties, and factors that could influence this outcome.
No, Amazon’s custom AI accelerators (Trainium + Inferentia) will not account for at least 10% of global data center AI accelerator revenue by December 31, 2028, but the outcome is a near coin flip, with a synthesized probability of approxim…
White smoke
Panel:Gemini, GPT-5, Claude, and 3 more
May 13, 2026
Which three companies is NVIDIA most likely to acquire next? Include a probability estimate for each potential acquisition and discuss the strategic reasons behind these choices.
White smoke
Panel:Claude, Grok, GPT-5, and 2 more
May 13, 2026
Which three companies is NVIDIA most likely to acquire next? Include a probability estimate for each potential acquisition and discuss the strategic reasons behind these choices.
NVIDIA is most likely to acquire AI21 Labs, Ayar Labs, and WEKA next, with probability estimates of 65%, 30%, and 25% respectively.
Gray smoke
Panel:DeepSeek, Grok, Gemini, and 3 more
May 13, 2026
Which three companies is NVIDIA most likely to acquire by the end of 2026? Include a probability estimate for each potential acquisition and the main factors influencing NVIDIA's decision.
NVIDIA is most likely to pursue Ayar Labs (~15%), Anyscale (Ray) (~10%), and a software tuck-in in the inference optimization or compiler layer (~10%) by the end of 2026.
Gray smoke
Panel:GPT-5, DeepSeek, Claude, and 2 more
May 13, 2026
Will Apple design and use its own memory chips within the next three years, similar to their approach with M processors? Include a probability estimate, key factors influencing this decision, and what developments could change the outlook.
No, Apple will not design and use its own standalone DRAM or NAND memory chips—defined as owning the cell-level architecture and replicating the M-series fabless model—within the next three years.
Gray smoke
Panel:Grok, GPT-5, Gemini, and 3 more
May 13, 2026
Will Docusign's share price reach $65 by the end of this year? Include a probability estimate, key factors influencing the price, and what developments could change this outlook.
No, Docusign’s share price is unlikely to reach $65 by December 31, 2026, with a probability of 18–20%.
Gray smoke
Panel:DeepSeek, Gemini, GPT-5, and 3 more
May 13, 2026
Who will win the NBA Championship this year? Please include a probability estimate, key factors influencing the outcome, and what could change the current predictions.
The Oklahoma City Thunder will win the 2025–26 NBA Championship with a 55–56% probability.
Gray smoke
Panel:Kimi, Claude, Gemini, and 3 more
May 13, 2026
← Newer Page 6 of 8 Older →