New AI Protocol Lets Models Debate Like Experts

Researchers unveiled the Consilium Protocol, a new way for AI models to debate like a team of experts. Instead of just picking the most confident answer, the system treats disagreements as useful signals. It assigns each AI a specific cognitive persona, like a doctor or engineer, to separate what a model knows from how it reasons.

The protocol is inspired by Byzantine Fault Tolerance (BFT), a concept from distributed computing that helps systems reach agreement even when some components are faulty or untrustworthy. Here, it's adapted to treat inter-model disagreement as epistemic signal rather than error. The system also introduces an In-Sample/Out-of-Sample validation framework borrowed from quantitative finance to distinguish conclusions that stem from training-data consensus from those that are empirically grounded.

This matters because it could make AI more trustworthy. Imagine asking a doctor, lawyer, and engineer the same question—you’d get different answers, but all useful. This system does that automatically, giving you the best of each perspective. It’s like having a roundtable of specialists at your fingertips, without the need to consult each one separately.

The protocol was tested across 1,478 deliberation sessions spanning 32 topics, demonstrating its ability to surface robust, well-reasoned conclusions from diverse model perspectives.

If you’re curious, you can read the full research paper on ArXiv. Just go to https://arxiv.org/abs/2606.00005 and search for the paper title. It’s free and open for anyone to explore.