Typed output: schemas, validation and tool calls

A model is an untrusted upstream service that happens to speak English. Hand its prose to a CRM and you have built a parser that fails silently. Structured outputs fix the shape of what comes back: JSON mode guarantees parseable JSON, schema-enforced modes guarantee the JSON matches your schema, and both work by constraining which tokens the model may sample. None of them guarantee the values are right. So you define the contract once in code (a Pydantic model), validate every response against it plus the business rules a schema cannot express, feed errors back for a bounded number of repairs, and route to a human when the budget runs out. Function calling is the same mechanism pointed at your code: the model proposes a call as structured arguments; your code validates, authorises and executes it, then returns the result.

• Shape is guaranteed, truth is not. Strict mode removes parse errors and missing keys, not wrong answers.
• The typed model is the source of truth. Generate the schema from it, validate against it, version it.
• Always give the model a legal "unknown". A schema with no escape hatch forces fabrication.
• Repair loops are bounded and end at a human queue, never at "write it anyway".
• The model proposes the call; your code decides whether it runs.