# Verification checks (P4) — 2026-10-30

Reproducibility: deleted out/results.json, out/analytic.csv, out/exclusions.csv; re-ran ~/wp29/venv/bin/python out/analysis.py; SHA-256 of all three files identical to the first run — PASS.

Internal consistency (all PASS):

• analytic rows == n_analysed (345); unique (iso3, year); no NaN; each balanced country has 23 years; country list matches balanced_countries.
• Exclusion ledger: n_before − n_removed = n_after each row; chained; starts at 439 = rows read; ends at 345 = n_analysed. Total removed 94 = 23 duplicates + 1 missing + 70 incomplete-country rows.
• n_countries_balanced (15) + n_countries_incomplete (5) = n_countries_total (20).
• change = mean_2022 − mean_2000 (identity holds to 1e-5).
• Independent recomputation via iso3×year pivot reproduces balanced set, both balanced means and the 2022 baseline (<1e-6).
• Magnitudes within 40–90 years; balanced 2022 mean (81.55) > all-available 2022 mean (78.69), as expected since the low-life-expectancy countries are excluded from the balanced set.
• Info: a naive mean over raw 2022 rows that double-counts TUR would be 78.639571 (vs 78.692 deduplicated) — confirms dedup matters for the baseline. No CI pairs, no chronology pairs, no random seed (deterministic).