Case Studies
Tested against real elections.
No synthetic data.
RPTK has been backtested on real campaigns using only real articles and no retrospective fitting. Every prediction is auditable.
NC Governor 2024
Robinson vs. Stein. 39 real articles scraped from AP, CNN, NPR, and WRAL. The model detected a scandal-driven collapse and called the winner within 3.2 points — with no polling data.
Called Josh Stein as winner before election day
Predicted 57.7% Stein vs. 55.5% actual — 2.2-pt error
Detected CNN "Black Nazi" story as the decisive inflection point
Conservative personas held a floor, preventing unrealistic collapse
3.2pt
Margin error
57.7%
Predicted Stein
55.5%
Actual Stein
39
Real articles
0
Polls used
UNC Student Body President 2026
Three candidates, a mid-race scandal, and 30 articles. Called the winner, the runner-up, and the dropout in 6 minutes for $2.43.
Winner
Duncan
2nd
Ahmad
Out
Bala
Detected scandal as the top inflection point
Predicted withdrawal before it happened
Cost under $3 total
Live Campaign · 2026
Currently deployed on a live North Carolina campaign with real-time monitoring, messaging optimization, and persona-tested content.
Full candidate profile and opposition dossier
Sentinel social media monitoring (Instagram, local news)
All content drafted and tested through voter personas
Backtested on 2024 cycle data to validate the model
86%
Focus group match rate
3.2pt
Best margin error
100%
Winner accuracy
Every backtest processes articles in strict chronological order — no future information leaks into past predictions. Every persona reaction, approval score, and vote-share calculation is logged and traceable.