Recall is 2 points short. One team wants to double npro...

Doubling nprobe doubles the scan to 48,828 candidates, which stays inside the block-structured kernel and has no new failure modes. Exact reranking requires 200 scattered random reads; while the volume is small, the cost is dominated by access latency rather than bytes, reintroducing the divergent access pattern the index was designed to eliminate. The staff-level decision depends on whether the full vectors are resident in RAM; if they are, reranking is nearly free, but if they are on disk, the latency cost is likely prohibitive.