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