The Synthetic Study

The 1,002 personas were generated by Google's Gemini 2.5 Flash, stratified across ten regions: Western Europe, Eastern Europe and Central Asia, East Asia, Latin America, the Middle East and North Africa, sub-Saharan Africa, North America, South and Southeast Asia, Oceania small states, and a transnational diaspora category. Each persona came with a biographical narrative, demographic attributes (age, gender, occupation, education, economic position, religious tradition, governance experience), and a statement of the tensions the persona carries.

The instrument was administered twice to 150 personas — once by Claude Sonnet 4.6, once by Gemini 2.5 Flash — and once to the remaining 852, split evenly between the two models. That produced 1,152 administrations across 1,002 personas: 576 by Claude, 576 by Gemini, 150 shared.

Scored profiles were clustered in the 12-dimensional axis space using k-means. Six clusters emerged as the silhouette peak across k=6 through k=18. Those six clusters were then compared against the twelve hand-crafted archetypes; the comparison informed the archetype revision documented on the Archetypes page.

The clusters are empirical — they emerge from the scored axis profiles, not from theoretical archetype definitions. Which hand-crafted archetypes survive contact with the data, and which get revised, can be asked and answered.

The 150 shared personas support model-level comparison: where Claude and Gemini agreed, where they diverged, and whether disagreement correlates with persona attributes.

The tensions — places where a persona's forced-choice, scaled, and budget responses pull in different directions — can be located axis by axis and cluster by cluster. That tells us something about which axes the instrument measures cleanly and which create internal conflict in respondents.

These personas are synthetic. The regional and demographic distributions reflect how Gemini was prompted to generate them, not the actual distribution of humans in those regions. The cluster shares (16% for C0, 21% for C2, and so on) say something about the shape of the 12-axis space as traversed by these specific personas — not about the prevalence of any governance philosophy in the real world.

Per-country analysis runs into the same limit: most countries have too few personas to support inference. Regional aggregates are on firmer ground, but “firmer” here means descriptive of this dataset, not extrapolable to the populations behind it.

The study can tell us about the instrument. It cannot tell us how real populations would answer it.