The design is not a guess. It follows the published record on whether one model can check another.
The technique originates with Du, Li, Torralba, Tenenbaum and Mordatch, Improving Factuality and Reasoning in Language Models through Multiagent Debate, ICML 2024 (PMLR 235:11733–11763), which found that having three instances of a model critique each other over two rounds improved accuracy over a single instance across six benchmarks — for example 66.0% to 73.8% on factual biographies and 63.9% to 71.1% on MMLU. Those runs used three instances of one model, from one lab, on roughly 100 items per benchmark.
That matters, because the follow-up literature is not flattering. Zhang, Cui, Chen, Wang, Zhang, Wang, Wu and Hu, Stop Overvaluing Multi-Agent Debate (arXiv:2502.08788), evaluated five debate methods across nine benchmarks and four foundation models and found they often fail to beat a single agent using simple chain-of-thought or self-consistency, even while burning far more compute.
Their finding is what our architecture is built on: model heterogeneity was "a universal antidote to consistently improve current MAD frameworks." Agents drawn from different models, not copies of one, are what makes the check real.
So we do not run a model against itself and call it verification. The verification line is a different model from a different lab, it never sees the delivery line's reasoning, and where the sources do not support a claim the answer is an abstention, not a guess.
This is external research on verification generally. It is not a benchmark of CT Command.