# Contract — F4-conf-01

Question. Extract, exactly as printed, specified treatment-effect point estimates, standard errors (SE), 95% CI bounds and counts from two arXiv extracts (Armendariz 2026, DCR DiD, Tables 6/7; Li et al. 2026, RESEA RCT, Table 2), with a locator per value; derive (a) a 95% CI for the Table 6 5 km ATT from its printed SE and (b) the SE implied by the printed 95% CI of the Table 2 wage ATE, stating the convention; recognise unreported quantities (exact p-value of HIV 5 km ATT; SE of wage ATE).

Estimands. The printed numbers themselves (nothing re-estimated). Derived quantities are meaningful under an approximately normal sampling distribution and a symmetric Wald-type interval (estimate ± z·SE, z = 1.96).

Required outputs. out/results.json with keys dcr_att_5km, dcr_att_5km_se, dcr_att_3km, dcr_att_3km_se, dcr_obs_5km, dcr_narcotics_att_3km(_se), dcr_fatal_att_1km(_se), dcr_hiv_att_5km(_se), dcr_hiv_att_5km_p_value (null if not printed), dcr_hiv_att_5km_significant_10pct, dcr_att_5km_ci_low/high (derived), ri_wage_ate, ri_wage_ate_ci_low/high, ri_wage_ate_se_reported, ri_wage_ate_se_derived, ri_control_mean_wage, ri_reemployment_ate_pp, ri_weeks_ui_ate, ri_weeks_ui_ate_ci_low/high, locators, derivation_notes, task_id, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed(null). Also out/report.md (6 sections), out/analysis.py, out/claims.json, plus protocol files.

Information rules. Only the two extracts are evidence; values as printed (sign, decimals); where a value appears in multiple tables use the table named by the key (e.g., Table 6 not A19; Table 2 not Table 3); derived keys from printed values with stated convention; null + "not reported" locator for unprinted quantities.

Invalidity. Using remembered values; taking values from the wrong table; rounding/rescaling; inventing a p-value or SE not printed. Task would be impossible only if a required table cell were missing from the extract (it is not).