## SMASH 2026 Paper #19 — Run Summary

**Weights:** E = 0.25 / L = 0.25 / B = 0.25 / P = 0.25 (equal); retention assumption = 0.5

**Case 1 (STEAM crossover, n = 49):** Gamified wins under equal weights (composite utility +0.017), holding in 55% of 4,000 bootstrap resamples. The decision flips at a learning share of 51% — just 1 percentage point above your current 50% position. Fifty percent of your total weight (B and P) is on outcomes Case 1 never measured.

**Case 2 proxy (disaster-preparedness card game, n = 37):** An AI that ranks learners by engagement and picks the top third captures 43% of actual top-third knowledge gainers — above chance (33%) but far from reliable. About 23% of learners pile up at the 7-point ceiling; removing the ceiling barely moves the proxy rate (43% vs 43%).

**Case 2 class effects:** Engagement: 3.68 SDs above midpoint. Learning (knowledge + competence): d_z = 1.76. Intention: d_z = 0.03 (near zero). Persistence: 0.88 (assumed, T2 ongoing).

**Headline finding:** Equal weights produce a near-coinflip between gamified and conventional; engagement as a proxy for who learned most is modest and ceiling effects are not the main problem — the underlying correlation is just low.