We use three classes of signals: * Tool-calling success and reliability from real traffic * Provider performance metrics such as throughput and latency * Benchmark and evaluation data as it becomes available
