You seem to have missed the calibration steps and the entire point of this work, Jim, but thanks for the comment!
The method was validated against real human survey data :57 products, 150-400 real respondents each. Synthetic responses hit 90% of the human-to-human agreement ceiling, and reproduced demographic patterns (income sensitivity, age effects) that weren't programmed in.
The paper and code are public if you'd like to dig into the methodology (links in post).
You seem to have missed the calibration steps and the entire point of this work, Jim, but thanks for the comment!
The method was validated against real human survey data :57 products, 150-400 real respondents each. Synthetic responses hit 90% of the human-to-human agreement ceiling, and reproduced demographic patterns (income sensitivity, age effects) that weren't programmed in.
The paper and code are public if you'd like to dig into the methodology (links in post).