Researchers just showed that the names you pick for JSON output fields can change how well a language model solves math problems - no prompt edits required.
The study tested seven language models on the GSM8K and Math500 math benchmarks using constrained decoding, a technique that forces a model to output valid JSON matching a schema. The researchers varied only the wording of the schema keys - the field names inside that JSON - while leaving the prompt untouched. Accuracy moved as a result, sometimes up and sometimes down, and the direction depended on the model being tested. The researchers also found that instructions placed in the prompt and instructions embedded in schema-key wording do not simply add together; combining them can produce effects neither one produces alone. The pattern was much clearer on GSM8K than on Math500.
Most engineering teams treat a JSON schema as plumbing: a contract for parsing output, not a lever on reasoning. This research says otherwise. Two teams could ship functionally identical schemas with differently worded field names and get measurably different accuracy, with neither team aware the field names were doing any work at all.
It's a familiar shape of problem dressed up in a new place. Prompt wording has been known to be brittle for years; now the schema you bolted on to tame that brittleness turns out to carry the same risk. "Just add structured output" was never going to be a free lunch.