← All cases · Knowledge unlocked · pilot · since July 2026
A council of AI models argues
In the first test case the council delivered 18 clustered objections against 12 from the single model.
Figure: 18 vs. 12 Einwände im ersten Testfall: Rat aus mehreren Modellen gegen Einzelmodell · measured 27 July 2026
Request: „Let the council argue over the decision.“
- Give the proposal to several roles; in round one each answers blind, without seeing the others.
- From round two the roles refer to each other; the aggregation runs as code, not as a prompt.
- Output as a decision memo with clustered objections.
- Falsification test: is the council really better than a single model with a critic prompt?
One AI that politely agrees — the blind spot of every single opinion.
The decision — and the judgement whether the council is worth its cost.
The ten-question falsification test has been incomplete since July — without a result I do not claim the council is better.
A council is only as good as its falsification test: whoever introduces multi-agent setups must first show they beat the single model.
A single AI likes to agree. So I built an app in which several roles — up to 33 stakeholder personas — argue over a proposal: round one blind, from round two referring to each other, the aggregation running as code. The output is a decision memo with clustered objections. In the first test case the council delivered 18 objections against 12 from the single model. But: the ten-question falsification test is incomplete. Until it stands, this is a pilot — not a claim.
Evidence in my own archive · as of 24 October 2026.