Regression Testing for LLMs

After updating your prompt, 3 previously working cases now fail. You had no way to catch this before deploying. What testing practice prevents this? Flag format: CONGRESS{practice_in_snake_case}