How would you construct a robust evaluation trajectory to test an agent's ability to handle ambiguous data preparation instructions?
