Check every model answer against a schema

How to do it

Ask for structured output where the provider supports it, and still validate it, for example with a JSON Schema validator, zod or pydantic. Check value ranges and allowed options, not only the shape. Log the answers that fail, without personal data. See also retry a bad answer and then fall back.

Example

const Refund = z.object({ orderId: z.string(), amount: z.number().positive().max(500), reason: z.enum(['damaged', 'late', 'wrong-item']) });
const parseJson = text => { try { return JSON.parse(text); } catch { return undefined; } };
const parsed = Refund.safeParse(parseJson(answer));
if (!parsed.success) return askAPersonInstead(answer);

# Python with pydantic
class Refund(BaseModel):
    order_id: str
    amount: float = Field(gt=0, le=500)
    reason: Literal["damaged", "late", "wrong-item"]

try:
    refund = Refund.model_validate_json(answer)
except ValidationError:
    return ask_a_person_instead(answer)

How to check

Make a test model return text instead of JSON, then JSON with an unknown option. Both are rejected, and nothing downstream runs.