1[2 {"id":"C1","text":"The unadjusted association between training hours and next-year productivity growth is negative (-0.132 pp/hour, 95% CI excludes zero).","results_keys":["naive_estimate","naive_ci_low","naive_ci_high"],"status":"supported","caveat":"Association only; confounded by skill_index."},3 {"id":"C2","text":"Adjusting for pre-programme skill_index, one extra training hour raises productivity growth by about 0.254 pp, 95% CI [0.234, 0.273]; conclusion code positive_effect.","results_keys":["effect_estimate","effect_se","effect_ci_low","effect_ci_high","conclusion"],"status":"supported","caveat":"Causal only under selection on observables with skill_index as the sole confounder, linear homogeneous effect; untestable with these data."},4 {"id":"C3","text":"The sign reversal is explained by confounding: skill_index is strongly negatively related to growth and positively to training.","results_keys":["skill_index_coefficient","naive_estimate","effect_estimate"],"status":"supported","caveat":"OVB identity verified in-sample; does not rule out other confounders."},5 {"id":"C4","text":"The adjusted model fits much better than the unadjusted one (RMSE 1.98 vs 3.86 pp).","results_keys":["model_metric","baseline_metric","adjusted_r2","naive_r2"],"status":"supported","caveat":"In-sample fit; better fit does not establish causality."},6 {"id":"C5","text":"The effect estimate is robust to a cubic skill adjustment and to including the 12 negative-hour rows.","results_keys":["sensitivity_cubic_skill_estimate","sensitivity_fullsample_effect_estimate","effect_ci_low","effect_ci_high","n_excluded_negative_hours"],"status":"supported","caveat":"Robustness to unobserved confounding was not evaluated."}7]8