Checking happens on this device. Your schema and answers are never uploaded.
Why check answers in bulk
Models asked for structured output usually get it right, but not always: a missing field, a number sent as text, a value outside the allowed list, or JSON wrapped in a sentence. Running a batch of real answers through your schema shows how often that happens and which rules trip the model up, before the problem reaches your users.
Pasting answers
Paste one answer per line, a JSON list of answers, answers separated by lines of three dashes, or a chat log with JSON in code fences. Open several .json files at once and each file counts as one answer; a .jsonl file counts as one answer per line. With code fences and text accepted, JSON is taken from inside answers like "Sure! Here it is: ...".
Questions
Which schema features are supported?
Types, required and optional properties, additionalProperties, enum and const, string length, pattern and common formats, number ranges, array items, length and uniqueness, allOf, anyOf, oneOf, not, if and then, and local $ref definitions. That covers the schemas used for structured outputs and tool calls.
Can I get a schema without writing one?
Make a schema from the first answer builds one from a known good answer. Tighten it afterwards, for example by adding allowed values or removing fields that are optional.
What is in the report?
The CSV has one row per problem with the answer number, its line, the path inside the answer and what's wrong, plus a row for each answer that passed. Download passing answers saves just the valid ones as JSONL.