The models aren’t static, we have to build validation sets to measure model drift and modify our prompts to compensate.