Structured JSON Output + Validation

The moment an LLM feeds a downstream system, plain prose is a liability. You ask the model for JSON, and then you must never trust it blindly. Models emit malformed JSON, drop fields, and put the wrong type in a slot. A production parser assumes all of that will happen and checks every field.

You will build the validator that guards the boundary. validate_record(text) takes a raw string the model returned and:

• Tries to json.loads it inside a try/except, so malformed JSON returns a clean failure instead of crashing your service.
• Checks the required schema {"name": str, "age": int, "email": str}: every field present, every value the right type.
• Returns {"ok": bool, "errors": [...]} where errors names what went wrong.

There is one trap that catches almost everyone. In Python, bool is a subclass of int, so isinstance(True, int) is True. JSON true parses to Python True, and a naive type check would accept "age": true as a valid integer. Reject booleans where you require an int.

>>> isinstance(True, int)
True            # the trap: True sneaks past an int check
>>> type(True) is int
False           # one way to be strict

A model could fill this JSON for you live, but grading only touches your deterministic validator: does it accept a clean record, and does it reject every flavor of broken one? Press Run to grade.