“Research shows that developers' criteria for 'good' and 'bad' LLM outputs evolve as they review more examples, a phenomenon known as 'criteria drift', making it impossible to define a complete evaluation rubric upfront.”