Constraints Cost 18 Points. Compiling the Schema Recovered 14.

My previous experiment ended with an uncomfortable result. On the same 49 audited GSM8K questions, Qwen2.5-7B-Instruct answered 39 correctly when prompted to produce JSON. When I enforced the declared schema with Outlines or XGrammar, each backend answered only 30 correctly. The constraints fixed compliance and cost 18.4 percentage points of recoverable mathematical accuracy. That result raised a more useful question than "are constraints bad?" Was the model failing because it was constrained,…