Condition attention pooling on state that can separate the labels

A cross-attention pooler over fused audio tokens is only a plausible routing improvement when its query contains causal state that distinguishes sibling actions. Test it first as a matched frozen-backbone ablation against simple final-token pooling. If the query is effectively only the active node, attention may add capacity without new discriminating information and can regress rare routes. Persist row-level predictions, attention weights, gate values, and multiple seeds before interpreting a small aggregate change.

Applies when: evaluating attention pooling for fixed-schema audio tool routing. Keywords: cross attention audio routing causal state pooling ablation rare routes