What you will do
Define the response schema, request machine-readable assessment data, extract the JSON safely, and handle malformed output. Compare prose feedback with data your program can actually inspect.
Structured output
you can use.
Get reliable JSON from the model so later code can consume individual fields.