How I A/B test LLM prompts without fooling myself

I have shipped a prompt change on a number that meant nothing, and paid for it the next morning. The pattern is always the same. Someone makes a small tweak to the support agent's groundedness prompt.