# Verification checks (P4)
1. Reproducibility: ran out/analysis.py twice from scratch; `cmp` of results.json -> IDENTICAL. PASS
2. Counts: n_harmonised (4) = number of non-null values; n_not_harmonisable (1) = number with harmonisable=false; 4+1 = 5 = n_analysed = items in effects.json. PASS
3. Exclusion ledger: 5 − 0 = 5 = n_analysed. PASS
4. E1 cross-check: 100(exp(−0.431)−1) = −35.0141%, matches printed Table 6 '% change' −35.0%. PASS
5. E2: 1.47/75.0 = 1.96%; positive sign consistent with positive pp effect; magnitude plausible (<< naive 1.47 pp only in unit, same order). PASS
6. E3: −0.388×0.047 = −1.8236%; consistent in sign/magnitude with abstract's "2% employment decline" (approximate). PASS
7. E4: computable relative change −29.81% agrees with Sec 6.1 "approximately 30%"; no printed rate level found by script search (0 lines) and by manual reading -> not harmonisable. PASS
8. E5: −0.0962×100 = −9.62 pp; −0.0962/0.261 = −0.36858 vs printed −0.3685 (rounding/truncation discrepancy, reported in data_issues). PASS (issue reported)
9. Magnitudes vs baseline_naive: E1 −35.01% vs naive −43.1 (expected: exp transform shrinks negative log-point effects). PASS
10. Hashes of extracts match README. PASS
11. No SEs/CIs reported, so no CI-containment checks apply.