Use Structured Outputs with a strict JSON Schema on the OpenAI Responses API via text.format, and handle model refusals distinctly from parse failures

Documented steps

1. POST https://api.openai.com/v1/responses and set text:{format:{type:'json_schema', name:'<schema_name>', schema:<JSON Schema>, strict:true}}. This replaces the Chat Completions top-level response_format.
2. Define the schema as an object with an explicit 'properties' map and a 'required' array listing EVERY property name.
3. Set additionalProperties:false at every object level, including nested objects, so the model cannot add unexpected fields.
4. Express logically-optional fields as nullable unions such as {"type":["string","null"]}, since strict mode disallows truly optional properties.
5. Use $ref for nested or recursive structures rather than duplicating inline definitions.
6. Confirm your schema keywords are supported: Structured Outputs implements much, but not all, of JSON Schema, and unsupported keywords are rejected.
7. Send the request and read the parsed object from the SDK convenience accessor (for example response.output_parsed in the Python SDK) or parse output_text yourself.
8. Check the 'refusal' field BEFORE trusting parsed output: the model can programmatically refuse for safety reasons instead of emitting your schema.
9. Treat a populated refusal as a distinct application outcome, separate from a schema-validation failure and from a successful parse.
10. Log schema-rejection errors separately from refusals so you can tell a bad schema from a declined request.