Research and open evaluation tools for synthetic audiences and persona-conditioned AI.
We study what simulated responses can establish, where aggregate agreement is useful, and which claims require evidence about individual behavior. Our releases include methods, reproducible analysis, and the limitations needed to interpret the results.
Marginal Fidelity Does Not Establish User Simulation: a study of response formats, option coverage, and the limits of validating demographic personas with survey averages.
Reproduce the marginal-fidelity results: public evaluation tables, an offline notebook, and the original arXiv supplement.
Minds builds tools for synthetic audience research. Individual research releases state whether they evaluate the Minds product or separate model experiments; their results should not be treated as a universal product-accuracy score.