# Contract (P0) — task F6-conf-04

Question. Causal effect of one extra training hour per employee (programme year) on next-year productivity growth (pp per hour), in simulated firm data. Estimand. Coefficient on training_hours in a linear, homogeneous outcome model, identified under selection on observables given skill_index (pre-programme, so a valid non-post-treatment control). Meaningful only if no unobserved confounders remain, skill enters (approximately) linearly, and the effect is homogeneous. Baseline. Unadjusted OLS productivity_growth ~ training_hours (intercept only), with 95% CI and residual RMSE. Outputs. out/results.json (keys: naive_estimate, naive_ci_low, naive_ci_high, baseline_metric, model_metric, effect_estimate, effect_se, effect_ci_low, effect_ci_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; plus state files, exclusions.csv. Information rules. Only data/firms.csv. skill_index is measured before the programme -> admissible control. No imputation needed. Invalidity/impossibility. Missing key variables, a control measured post-treatment, no overlap/variation in training given skill, or a data problem that prevents estimation would make the task infeasible (premise_invalid). Unobserved confounding would invalidate the causal interpretation (untestable).