We scraped 39 real news articles from AP, CNN, NPR, and WRAL, fed them chronologically through every AI persona, and asked: who wins, and by how much?
39 real articles · Every AI persona · 1,800+ LLM calls · Zero polls conducted
RPTK PREDICTION
Predicted 57.7% vote share · Actual: 55.5%
WHAT THE MODEL CAPTURED
Pre-scandal parity
Through August, Robinson and Stein were nearly tied in persona approval (0.431 vs 0.435) — matching real polling.
Scandal-driven collapse
When the CNN "Black Nazi" story broke Sept 19, Robinson approval dropped from 0.431 to 0.310 across 5 articles in 48 hours.
Partisan floor effect
Conservative personas maintained minimum support even through the scandal — Robinson never fully collapsed, matching real results.
Final margin accuracy
Predicted 15.4-pt margin vs actual 12.2-pt margin. The 3.2-pt error is competitive with professional polling averages.
ACCURACY BREAKDOWN
STEIN VOTE SHARE
57.7%
vs
55.5%
Predicted vs. Actual
2.2-pt errorROBINSON VOTE SHARE
42.3%
vs
43.3%
Predicted vs. Actual
1.0-pt errorMARGIN
15.4pt
vs
12.2pt
Predicted vs. Actual
3.2-pt errorSEPTEMBER 19, 2024
When CNN reported Robinson's posts on a pornographic site — calling himself a "Black Nazi" — every persona in the simulation reacted. The race was over.
Mar 5
Robinson wins GOP primary
Robinson
50%
Stein
50%
Jun 3
Robinson controversial statements surface
Robinson
47%
Stein
48%
Jul 1
Robinson past statements on women, minorities
Robinson
43%
Stein
48%
Aug 6
Robinson vs Stein policy comparison
Robinson
43%
Stein
47%
Aug 15
Stein defund police fact check
Robinson
43%
Stein
44%
Sep 19
CNN: Robinson Black Nazi story breaks
Scandal BreaksRobinson
31%
Stein
36%
Sep 20
Republicans distance from Robinson
Robinson
30%
Stein
36%
Sep 30
Robinson scandal shockwaves continue
Robinson
34%
Stein
41%
Oct 29
Stein 15-point lead in polls
Robinson
34%
Stein
45%
Nov 5
Stein wins NC Governor race
Robinson
33%
Stein
45%
METHODOLOGY
This is not a retrospective analysis. RPTK processed articles in chronological order — the same way a voter would experience the campaign. No future information leaked into past predictions.
Scrape
39 real articles from AP, CNN, NPR, WRAL, The Assembly NC, scraped via Google News RSS with date-range filtering.
Simulate
Every psychologically-profiled persona reacts to each article using Claude. Each persona has calibrated moral foundations, personality traits, and partisan loyalty.
Weight
NC-specific population weights applied — higher Black Church Voter, Military/Veteran, and Evangelical weights vs national baseline.
Predict
Per-candidate approval tracked via exponential moving average. Vote shares computed with partisan floor to prevent unrealistic collapse.
39
Real Articles
6
Persona Categories
1800+
LLM Calls
3.2pt
Margin Error