Verification checks (P4)
- Clean re-run of out/analysis.py (results.json deleted, regenerated): byte-identical (cmp) — PASS
- counts sum to 6: PASS
- verdict labels valid: PASS
- keys A1-A6 in verdicts/locators/reasons: PASS
- locators start with cited source id or 'not reported in': PASS
- A1 computed pct rounds to printed -35.0: PASS
- A3 exp conversion: computed -24.648% vs printed -24.7% — FAIL at 1-decimal equality (computed rounds to -24.6). Handled: recorded as a source rounding data_issue (0.05 pp, likely unrounded coefficient); immaterial to the A3 verdict since asserted value is -28.3%. Reason text now states both values. Within 0.1 pp tolerance: PASS
- A4 CI contains estimate: PASS
- A4 verdict invariant to z=1.96 vs 2.0: PASS
- A6 rel pct: PASS
- exclusion ledger: 6-0=6: PASS
- Baseline vs final: naive verbatim-number baseline flags A3, A4 as mismatches (consistent with not_supported), but marks A2 (sign error: -0.063 vs printed 0.063 matches as digits) and A5 (no numbers) as matching — showing the baseline is insufficient; differences explained in report.
- Information rule: only data/ extracts used; script reads only data/ — PASS
- Re-run after edit: byte-identical — PASS