Core concept: A required field must be filled. If the source document has no value for it, the model has to invent one to satisfy the schema. Marking the field optional or nullable gives the model a legal way to say "not present."

Why A and B are correct:

The schema guidance is explicit: "Match schema to task. If the task might not have all the information your schema requires, make those fields optional."
The extraction example marks author and date optional "since git blame information might not be available for all files. The agent fills in what it can find and omits the rest."
Structured outputs are enforced by constrained decoding, with "Guaranteed field types and required fields." A required string gets produced no matter what the document says. So an over-strict schema is a hallucination generator.
A nullable type keeps the output structurally valid while carrying an honest empty value downstream. That is B: compliance without guessing.

Why the others are wrong:

C: Nullability is not a retry trigger. Retries fire on schema validation failure, and a null in a nullable field is valid output, so nothing retries.
D: tool_choice: "any" only forces the model to call one of your tools. Field nullability has no effect on it and cannot bypass it.

Remember: Optional and nullable fields let Claude report "missing" instead of guessing.

Question 170