We stopped asking an LLM how much its own work would cost

How CloudZero’s model-routing agent stopped asking an LLM to guess a token budget and switched to six bounded signals plus a deterministic formula: why averaging capped accuracy, why some signals dilute the ensemble, and where the v1.0.0 formula still needs calibration. The post We stopped asking an LLM how much its own work would cost appeared first on CloudZero .