Structured Output — JSON Schema, Pydantic, Zod, Constrained Decoding

"Ask the model nicely to return JSON" fails 5 to 15 percent of the time, even on frontier models. Structured outputs close that gap with constrained decoding: the model is literally prevented from emitting a token that would violate the schema. OpenAI's strict mode, Anthropic's schema-typed tool use, Gemini's `responseSchema`, Pydantic AI's `output_type`, and Zod's `.parse` are five surface forms of the same idea. This lesson builds the schema validator and the strict-mode contract learners will use for every production extraction pipeline.