Been shipping structured JSON output from LLMs in production for a health app. Here's what I've learned about reliability. The problem: get a 70B model to return valid JSON matching a strict schema, every time. What I tried: Attempt 1: "Return JSON." No schema. 40% valid output. Attempt 2: Detailed schema in prompt. 75% valid. Attempt 3: JSON mode enabled (Groq/OpenAI/Anthropic all support). 92%. Attempt 4: JSON mode + schema validator + retry loop with error surfaced back....