If you generate structured output from an LLM and validate against a schema, you know the failure mode: usually fine, occasionally a missing field or an unparseable response. The common fix is a retry, but a plain retry is the same prompt at the same temperature, so you are just re-rolling. What wor
