Model performance measured on the distribution of prompts that users naturally generate through interaction with that specific model.