Schema Adherence — AI Glossary

Overview

Schema adherence is the share of responses that parse and validate against the target schema on the first attempt, without repair. It differs from "supports structured output": a model may advertise the feature yet still emit missing fields, wrong enums, invalid dates, or extra prose under real prompts and edge cases. It matters because downstream systems reject invalid payloads, and every repair costs a retry. Measure it on a representative set, count first-pass validity separately from post-repair validity, and treat validation as part of the pipeline rather than trust. Common mistakes: testing only the happy path, ignoring long or ambiguous inputs where structure degrades, and treating a passing sample of ten as a reliability estimate.