Yeah this is an interesting way to compose the ensemble though. Rather than averaging over sub-networks they're synthesizing the panel of responses:

A judge model reads every panel response and produces structured analysis: consensus points, contradictions, partial coverage, unique insights, blind spots

Whereas beam search injects variance by trialling candidate sentences and mixture of experts has competing sub-models here we're reconciling different ideas.

Notice though that they're always using Opus 4.8 as the judge so I think the claim (of surpassing frontier models) is over-inflated. I'd instead characterise it as outsourcing legwork to cheaper models.

I'm still optimistic about ensembles of smaller models though. You could imagine a specialist synthesis model or advantages from combining latent activations instead of text responses