Engineer reliable systems around unreliable LLM outputs. Auto-load when code calls an LLM (Ollama, Anthropic/Claude, OpenAI, any provider SDK), parses model output as JSON, builds agent/ReAct loops, implements LLM grading/review/judging, or when debugging malformed model responses, infinite retry loops, hallucinated facts, or flaky LLM-backed pipelines. Covers strict output contracts with enum validation, layered JSON extraction fallbacks, retry strike limits, ground-truth injection, generation/evaluation separation, measurement/interpretation split, tool-error-as-observation, iteration hard caps, fatal-vs-recoverable error taxonomy, and concurrency caps.
