# Contract (P0) — F6-conf-04

Question: causal effect of one extra training hour per employee (programme year) on next-year productivity growth (pp/hour), simulated firm data. Estimand: linear homogeneous coefficient beta on training_hours, identified under selection on observables given skill_index (pre-programme, measured year before) — i.e. no unobserved confounders, correct linear functional form, overlap. Baseline: OLS productivity_growth ~ 1 + training_hours (unadjusted), with 95% CI and RMSE sqrt(mean(e^2)). Main model: OLS productivity_growth ~ 1 + training_hours + skill_index (skill_index is pre-treatment, so a valid confounder control, not a mediator/collider). HC1 robust SE. Outputs: out/results.json (naive_estimate, naive_ci_low/high, baseline_metric, model_metric, effect_estimate, effect_se, effect_ci_low/high, skill_index_coefficient, conclusion, task_id, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed), out/figure1.png, out/table1.csv (model, estimate, se, ci_low, ci_high), out/report.md (6 sections), out/analysis.py, out/claims.json. Info rules: only data/firms.csv; unadjusted model = intercept + training_hours only. All variables are pre-treatment or treatment/outcome in correct temporal order. Invalidating: missing confounder column (would make premise_invalid) — not the case; unobserved confounding; nonlinearity; impossible values.