The per-language and per-transport point is exactly where a routing benchmark becomes useful instead of just another leaderboard. I’d keep the scorecard decompo

The per-language and per-transport point is exactly where a routing benchmark becomes useful instead of just another leaderboard. I’d keep the scorecard decomposed: transport/streaming latency, WER or task accuracy by language, TTS quality, and effective cost under a representative context/cache distribution. Then expose the constraints and raw measurements with the selected route. Otherwise a single composite score can hide a provider that wins English batch tests but fails the production transport or language.