Benjamin Nweke discovered a subtle pipeline failure after enabling Structured Outputs for parsing payment messages. His reconciliation job flagged mismatches where transaction dates were incorrect. Investigation revealed the schema declared the date as required, so the model returned the extraction run date instead of leaving the field blank. The amount and sender matched perfectly; only the date was fabricated.

Similarly, an on-premise retrieval system implemented a semantic cache using cosine similarity with a 0.92 threshold to serve stored answers and skip generation. However, the cache lacked a mechanism to return "no answer" when questions were near-identical but had opposite correct answers. This created a failure mode where the system confidently served an incorrect cached response, leaving no trace to distinguish a correct hit from an incorrect one.

Both components were designed to improve reliability by narrowing output possibilities. A schema narrows output to a fixed shape, while a threshold narrows a continuous score to a binary decision. However, this narrowing removes the degree of freedom for abstention. Constraint violations are loud—failing parsers or type checks—but a forced value passes every gate because the problem is provenance, not format. A required field filled leaves an inspectable record, whereas a cache hit leaves no generation to compare, allowing incorrect answers to reach users as fluent, sourced confidence.