Consilium Protocol: Cheap AI models rival frontier ones in structured debates

A new paper by VD Doske (arXiv:2606.00005) presents the Consilium Protocol, a Byzantine Fault Tolerance-derived architecture that transforms inter-model disagreement into structured epistemic signal rather than error. The protocol assigns engineered cognitive personas to language models, separating underlying architecture from reasoning behavior, and uses an In-Sample/Out-of-Sample validation framework borrowed from quantitative finance to distinguish training-data consensus from empirically grounded conclusions.

Over 1,478 deliberation sessions across 32 topics in 10 domains, the protocol revealed four startling findings: (1) the cognitive persona determines epistemic behavior—free edge-inference models costing just $0.0002 per batch matched the analytical output of frontier models at $10.69 per batch; (2) RLHF alignment training creates domain-specific blind spots—contested policy topics saw 12.3 percentage points less adversarial challenge than settled science topics, and AI safety discussions showed asymmetric bias where models challenged claims that 'AI is dangerous' far more vigorously; (3) the protocol itself exhibits no directional bias (immigration Δ=2.3%, renewables Δ=1.2%); and (4) out-of-sample evidence retrieval validated 239 claims with 100% success and uncovered 167 blind-spot discoveries invisible to training-data deliberation. Run-to-run reproducibility averaged ±2.2% standard deviation, and the authors released the protocol specification under MIT license.