Systematically investigate asymmetric consensus-label thresholds

Investigate systematically the effects of independently varying the outlier-eligibility threshold, positive-label threshold, and negative-label threshold in the consensus labeling scheme for anticipatory outliers.

Background

The study labels articles by requiring agreement across multiple embedding models about both publication-time outlier status and later anticipatory trajectories. Its experiments use the conservative symmetric setting $(k_o,k_a,k_n)=(k,k,0)$, in which the same threshold supports outlier eligibility and positive labels, while negative labels require no anticipatory votes. The paper notes that more general asymmetric threshold rules are possible, but their systematic calibration and evaluation are not undertaken.