PartitionLens: Identifying Trace-to-Route Alignment with Counterfactual Controls

Routed reasoning distillation faces an attribution challenge: an accuracy gain can reflect conditional capacity, a changed supervision pool, or trace-to-route grouping. PartitionLens resolves the comparison design by fixing the routed student, teacher-token budget, trainable-parameter budget, router depth, routing granularity, and paired evaluation pool, then using Latent-Routing MoA as the routed-capacity reference. On the reported five-domain primary pool, sequence-level PartitionLens-Open reaches $58.62$ A1, exceeding Latent-Routing MoA by $+2.46$ pp (paired $95$ BCa CI $[+1.76,+3.16]$). The advantage is positive in every reported domain, and its W/L/T accounting gives a $51.23$ tie-split win rate over $16,470$ paired outcomes. The result provides a precise matched-control reference for evaluating trace-grouped routed distillation.