Theoretically it should query a large number of models, and have a "confidence" based on how likely each model's answer is to be correct. Then return the most confidence answer.