Inspired by how bee swarms find the best nest - scouts argue for their picks, strongest conviction wins. Same here: models independently respond, then anonymously review each other (no playing favorites), vouch/challenge/build on ideas, and a chairman synthesizes the survivor.

The cross-pollination between different model biases is where the magic is. Gets outputs none of them produce alone.