Your digital twin is more predictable than you are: parsimony and dose-response analyses of the Twin-2K-500 digital twins

Using the public Twin-2K-500 dataset (2,058 US respondents, four survey waves, a digital twin for each respondent conditioned on about 500 of their answers), I ran two analyses in July 2026. Study 1 measures how many respondent variables are needed to predict 29 wave-four items and a purchase measure, for the real respondents and for their published twins. For real respondents, 14 census, political and religious variables capture 36 to 95% of what a 64-variable model reaches out of sample, and 39 variables reach at least 95% of it in five of six question families. For the twins, four variables (party, ideology, religion, religious attendance) predict political attitudes at R² = 0.72, against 0.24 for the real respondents; the twins' answers are more predictable from their profile (0.75) than respondents' answers are from their own earlier answer to the same item (0.54). Study 2 varies, for 100 respondents with engine and prompt template held constant, the number of the person's own earlier answers given to the model, from 0 to 500. No dose produces a detectable accuracy gain (0.673 with demographics only, 0.668 with 500 answers, change −0.005, 95% interval −0.028 to +0.019; test-retest ceiling 0.813). The accuracy result is consistent with Peng et al. (2026, Science Advances), whose preprint reported it first; the other four analyses do not appear in their paper in this form. The deposit includes the scripts, derived tables, the 600 prompts and model outputs of Study 2, and a manifest with checksums and the file dates recorded on the author's disk. Disclosure: the author's company, FlashInsight, sells synthetic samples.