Verification checks (P4)
Script: out/state/verify_checks.py; full output in out/state/verify_output.txt (32 PASS, 0 FAIL).
- Reproducibility: deleted results.json/table1.csv/figure1.png, re-ran analysis.py twice; results.json byte-identical (cmp), table1.csv and figure1.png identical md5. PASS. (Later re-run after a cosmetic figure-tick change: results.json unchanged.)
- table1.csv has 27 rows, 3 estimators; rel_bias_var empty for wcb. PASS
- coverage 27 keys, mc_se 27, rel_bias_var 18. PASS
- coverages in [0,1]; coverage×R integer; MC SE > 0 and equals sqrt(p(1−p)/R). PASS
- failure_cells/n_failure_cells and nominal_cells recomputed independently from coverage & MC SE: match. PASS
- R=2000, B=399, n_cells=9, n_analysed=18000 = 9×2000; random_seed = C1 seed 20261101. PASS
- Exclusion ledger: 18000 − 0 = 18000 = n_analysed. PASS
- rel_bias(CR1)/rel_bias(CR0) = G/(G−1) exactly in every cell. PASS
- Exact Var(ȳ) in each cell matches formula; empirical Var(ȳ) over R replications close to it (cell_diagnostics). PASS
- Analytic E[V̂_CR0]/Var(ȳ) (derived: Σ_g Var(T_g − n_g ȳ)/N²) vs MC estimate: |z| < 3 in all 9 cells; largest |z| = 2.46 (C2, G=5 ICC=0.2), others ≤ 1.40. With 9 comparisons one |z|≈2.5 is not unusual; PASS with note.
- CR1 mean half-width > CR0 in every cell (as must hold since t_{G−1}·sqrt(G/(G−1)) > 1.96). PASS
- Baseline (naive iid interval) behaves as theory predicts: rel. var bias ≈1.00 at ICC=0, ≈0.10 at ICC=0.2, ≈0.04 at ICC=0.5; coverage ≈0.95 at ICC=0 and 0.29–0.49 otherwise. PASS (magnitude sanity).
- Runtime ≈ 2 s ≪ 600 s limit. PASS