Repository navigation
First measured snapshot: 1,000 seeded replications, 12 periods, 5,000 reports and 200 gold labels per period.
- Classify-and-count turned a true 10% → 25% trend into 8% → 17% (slope × 0.628), and its naive intervals covered the truth 0% of the time.
- ACC and prediction-powered inference recovered the trend with ~95% coverage.
- A model upgrade read as a +24% "trend" under counting; PPI showed 0.0% change.
CI re-runs all 1,000 replications and fails if any published figure moves.