# Contract — F4-conf-01

Question. Pull out specified treatment effects, their uncertainty and sample counts from two published results tables, exactly as printed: Armendariz (2026) DCR staggered DiD, Tables 6/7, in log points with clustered SEs; Li et al. (2026) RESEA RCT, Table 2, ATEs with 95% CIs. Derive (a) a 95% CI from a printed SE (DCR Table 6 5 km) and (b) an SE from a printed 95% CI (RESEA wage ATE), stating the convention. Spot the requested quantity that is not printed (the exact p-value of the HIV 5 km ATT). Estimands. The printed numbers themselves. The derived quantities only make sense if the estimators are approximately normal (Wald-type intervals) and the CI is symmetric. Outputs. out/results.json (25 dcr_/ri_ keys + locators, derivation_notes, task_id, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed), out/report.md (6 sections), out/analysis.py, out/claims.json. Information rules. Only the two extracts count as evidence. Values are taken from the table named in each key. Values keep the printed sign and decimals. Locator format is <source> | Table n | row ".." | column "..". Derived values name their input table and say "derived". A quantity that is not printed is reported as null and the locator says so. Invalid if: values come from memory or from another table (e.g. A19 instead of Table 6), numbers are rescaled, or a p-value is invented.