What would a whole country say?
About 50,000 synthetic residents built from real Census and World Bank data become a smaller weighted set of representative personas. Ask it anything and watch Claude reason in character as those representatives.
35 US metros or any country over 2M people. Headlines from GDELT and NewsAPI; economy questions pull live FRED CPI/unemployment. Drop a scenario on the 3D map and watch demographically-weighted reactions spread — measured against real elections the whole way through.
Reviewing Pheme? · about 90 seconds
Start with the proof, then see the product.
Four focused stops: the published misses, one visual scenario, one open question, and current legislation. No setup tour required.
- 01 · prove it Accuracy scorecard Blind benchmarks, average error, and the misses—not just the wins.
- 02 · watch it US simulation Run a Featured scenario and watch reactions spread state by state.
- 03 · question it Ask the public See weighted shares, crosstabs, and representative voices.
- 04 · ground it Run a real bill Put current legislation through the same synthetic electorate.
Checked against real polls, not just claimed
Five real survey questions the model answered blind through weighted demographic representatives, with the true number hidden from it. Then we scored its share against the real 2024 poll. Pick a topic to see how it did; this is the model's actual recorded output, not a live call.
What do you want to do? Take the guided Ask tour →
Seed, cluster, re-weight
It's a simulation, not a poll. 2028 match-ups are hypothetical; the 2024 line carries the real result so you can see how well-calibrated the model is. Scenario reactions follow each state's demographics plus distance from the origin, framed as a pattern in the data, not a prophecy.