What it does
On 751 replayed routing decisions, Jev picked the right expert more often than the prompt-based router and answered 8.1× faster at the median. Reranking was the reality check: 40.2% Recall@6 against 44.3% for production, at roughly $0.00044 a query, with tie-breaking doing part of the work. Citation judgments repeated more consistently than GPT-5.6 Luna's. Their advice — benchmark it wherever the answers can be listed in advance — is the most reusable sentence in the piece.