How little instrumentation is enough?

Week 2, Lesson 2 addresses the question nobody asks explicitly but everybody is thinking: how many test queries are needed before confidence is sufficient to ship? This matters because every sample costs time and money. If a subject matter expert is writing oracle SQL and verifying outputs, that's real budget and real hours. But the alternative, shipping with insufficient evidence, means genuinely not knowing if an AI feature is at 75% quality or 85% quality. And that difference matters. 75% means one in four queries is wrong. 85% means roughly one in seven. Those are very different user experiences. Most teams guess. They say "let's run a hundred" or "fifty sounds good." No math behind it. Just vibes. This lesson replaces vibes with a formula. By the end, the minimum viable sample size for a specific decision at a required confidence level will be calculable. Exactly how much evidence is needed, and not a sample more.