Forcing structured output without breaking the model

JSON mode and schema constraints look like a free win until they aren’t. The first time the model produces valid JSON for a complicated nested schema, the temptation is to push everything through structured output. The second time it produces structurally valid but semantically empty fields, you start to see the trade-off.

Why constraints sometimes hurt

Forcing the model into a tight schema cuts off its room to reason. If your schema demands a single-paragraph summary and a binary classification, the model is going to commit to the classification before it has finished thinking about the summary. The order of fields matters. The required-vs-optional split matters. Required arrays of length 0 are the most common silent failure — the model satisfies the schema by giving you nothing.

Patterns that keep both

Two passes are simpler than they sound: a free-form reasoning pass, then a constrained-extraction pass that turns reasoning into JSON. You spend two model calls but stop fighting the model into a single one. For simpler tasks, ordered fields — reasoning first, decision last — do most of the work that two passes would.

The schema you ship is rarely the schema you sketched at the start.