How do you ensure consistency in test results generated via AI models?