A calibrated, falsifiable population-simulation forecaster for how public opinion shifts by political group after an event. Built on the real World Values Survey Wave 7 (97,220 respondents; US n=2,596), calibrated and validated on abortion / Dobbs.
Live: https://miroworld.renlab.ai
Given a real population and a policy shock, it predicts the by-party (Dem / Indep / Rep) move in opinion, broken down by group — the split a plain "ask an LLM" averages away.
- 3× more accurate than a vanilla LLM on a blind out-of-sample test (by-group MAE 0.030 vs 0.087).
- The partisan differential is recovered out-of-sample by two parameter-disjoint mechanisms (ideological-extremity weighting AND homophily social contagion) and by a real MiniMax LLM-agent run (48 calls: Dem +16 / Indep +11 / Rep ~0, diff +0.05 vs real +0.06) — evidence it's a real signal.
- Ground truth is real published polling (Gallup / Pew / Quinnipiac), every number re-verified at its primary source; unverifiable numbers are dropped, never fabricated.
policy-wind-tunnel/METHODOLOGY.md— per-agent update equations (extremity / contagion / LLM paths).policy-wind-tunnel/CROSSEVENT_FINDINGS.md— leave-one-event-out backtest, honest results + caveats.policy-wind-tunnel/CALIBRATION_FINDINGS.md— real MiniMax calibration on Dobbs.polymarket-backtest/ground_truth_events.md— the verified by-party ground-truth set (in the source tree).policy-wind-tunnel/PITCH.md— one-page overview.
policy-wind-tunnel/— Python package (pwt/), experiment scripts, findings, derived data.policy-wind-tunnel-site/— the deployed static site (globe + simulator).docs/— frontend source (pwt.js, overview + simulator HTML).
Calibrated and validated on abortion / Dobbs only; other issues on the site are illustrative. The absolute magnitude of a shock is exogenous (news salience) and reported as an interval, not a point estimate. Verified events are still few (n≈4) — the main thing to scale.