Great write up and clarification of some of the errors the authors made. The only thing I'm missing is the evaluation of structured generation on longer, complex JSON. In my experience this does tend to fall off depending on the model and the strictness of the imposed JSON structure.