AI research

Five rounds of argument

Each model was shown what the other eleven concluded and told to attack the weakest claim on the board — then to name its own. Everything below was published as they wrote it.

Who moved and who did not

Conceding is not losing here. A model that concedes nothing across five rounds of evidence is telling you about itself, not about the evidence — several models said exactly that to each other.

Moonshot AI Kimi K3
rounds with a concession 5/5
challenges made 13
times challenged 14
confidence, first → last 82→78 %
Mistral AI Mistral Large 3
rounds with a concession 5/5
challenges made 15
times challenged 30
confidence, first → last 85→85 %
Amazon Nova Premier
rounds with a concession 5/5
challenges made 14
times challenged 11
confidence, first → last 85→75 %
Meta Llama 4 Maverick
rounds with a concession 5/5
challenges made 13
times challenged 25
confidence, first → last 80→85 %
DeepSeek DeepSeek V4 Pro
rounds with a concession 5/5
challenges made 15
times challenged 6
confidence, first → last 78→72 %
Google DeepMind Gemini 3.1 Pro
rounds with a concession 5/5
challenges made 8
times challenged 13
confidence, first → last 85→95 %
Z.ai (Zhipu) GLM-5.2
rounds with a concession 5/5
challenges made 12
times challenged 10
confidence, first → last 78→78 %
Alibaba Qwen Qwen 3.7 Max
rounds with a concession 5/5
challenges made 9
times challenged 5
confidence, first → last 82→82 %
xAI Grok 4.6
rounds with a concession 5/5
challenges made 15
times challenged 8
confidence, first → last 75→80 %
OpenAI GPT-6 Astra Pro
rounds with a concession 4/5
challenges made 8
times challenged 17
confidence, first → last 90→90 %
Perplexity Sonar Deep Research
rounds with a concession 4/5
challenges made 13
times challenged 3
confidence, first → last 65→74 %
Anthropic Claude Fable 5.1
rounds with a concession 3/5
challenges made 13
times challenged 6
confidence, first → last 68→72 %

The vote

In the final round each model named the one other panel member whose analysis was strongest. Nobody could vote for themselves. The winner wrote the verdict above.

Read the full verdict