# Verification checks (P4)
| # | check | outcome |
|---|---|---|
| 1 | Re-run analysis.py from clean state (results.json deleted) → byte-identical results.json | PASS (cmp identical) |
| 2 | n_supported + n_not_supported + n_not_assessable = 6 = n_analysed | PASS (2+3+1=6) |
| 3 | exclusions: n_raw 6 − removed 0 = n_analysed 6 | PASS |
| 4 | Every verdict has one locator and one reason; locators follow the TASK format | PASS |
| 5 | Extract SHA-256 equal to README values | PASS |
| 6 | A1: 100(exp(-0.431)-1) = -35.01% matches printed -35.0%; |z| = 2.05 > 1.96; CI95 [-0.843,-0.019] contains estimate and excludes 0 | PASS |
| 7 | A3: 100(exp(-0.283)-1) = -24.65% (printed -24.7%) ≠ -28.3% | PASS (verdict not_supported) |
| 8 | A4: CI95 = [-0.896, 0.084] (1.96) / [-0.906, 0.094] (2.0); contains estimate; includes 0; asserted interval = b ± 1·SE | PASS (verdict not_supported) |
| 9 | A5: regex search for men/women/male/female/sex/gender in DCR extract returns no lines | PASS (not_assessable) |
| 10 | A6: -1.99/17.3 = -11.50%; CI in Table 2 = [-2.24, -1.75] | PASS |
| 11 | Baseline vs final: A3/A4 have only 75% of numbers found verbatim, both judged not_supported; A2 has 100% numeric match yet fails on sign — illustrates why verbatim match is insufficient | consistent |
| 12 | random_seed null (no randomness) | PASS |