When 100% of our prototype outputs were valid JSON but 0% met the data contract, we discovered the LLM was doing work that software should own. A field classification exercise and 4-pass pipeline brought schema compliance from 0% to 100%.