Your team ships caption-based unified fusion to control...

This is a multi-granularity noise-correspondence problem. Replacing captioning with raw-patch attention everywhere is inefficient, as it increases fusion compute by 5–55x for all traffic. Instead, implement a query-complexity router: use captions for compositional queries and reserve the expensive raw attention fusion for detail-specific queries where precision is critical.