# Verification checks (P4)

Clean re-run (deleted outputs, ran twice): results.json byte-identical -> PASS
numpy lstsq reproduces effect_estimate -> PASS
numpy lstsq reproduces skill coef -> PASS
model_metric = sqrt(mean e^2) -> PASS
CIs contain estimates -> PASS
CI = est +/- 1.96 se -> PASS
SE positive; RMSEs positive -> PASS
model_metric <= baseline_metric (nested OLS same sample) -> PASS
1200 - removed = n_analysed -> PASS
conclusion consistent with CI -> PASS
table1 main row == results -> PASS
omitted variable bias sign: skill coef<0 and corr(skill,hours)>0 imply naive<adjusted -> PASS
OVB identity naive = b1 + b2*delta -> PASS

spb/wp29_uqo