I’m considering DeepSeek for a pipeline that has to inspect a short record, make a judgment, and return only JSON for the next service. The tricky part is that the model needs to reason about the input, but the output parser rejects any explanation or formatting outside the JSON object.

How well does DeepSeek handle reasoning before emitting JSON in practice? I’d like to hear whether you’ve seen it add stray text, omit fields, or produce malformed output, and which model or prompting setup you tested.