Tag
This preprint challenges aggregate independence metrics for LLM judge panels, showing that verification signals only improve accuracy on pivotal 'one-vote-margin' queries, and proposing a margin-stratified call-reduction rule.