1# Verification checks (P4)21. Determinism: deleted out/results.json, re-ran analysis.py; `cmp` with previous output -> IDENTICAL. PASS32. Per-replication correctness: replication 1 slope, se_conv, se_HC1 match statsmodels OLS / HC1 to ~1e-15. PASS43. Exact SD: closed form sqrt(sum w_i^2 sigma_i^2)=0.267981 equals sandwich (X'X)^-1X'ΣX(X'X)^-1 [1,1]. PASS54. MC SD vs exact: 0.26965 vs 0.26798, diff 0.0017 < 1 SE of sample SD (~0.268/sqrt(2·1999)=0.0042). PASS65. Unbiasedness: mean b̂ 0.4962 vs 0.5, diff −0.0038 < 1 MC SE (0.27/sqrt(2000)=0.006). PASS76. E[se_conv]: MC mean 0.19026 ≤ Jensen upper bound sqrt(E[s²]·[(X'X)^-1]_22)=0.19318 (E[s²]=tr(MΣ)/28 analytic). Consistent. PASS87. Ranges: coverages in [0,1]; MC SEs >0; n_replications=2000, n_obs=30, n_input_rows=30; no exclusions (30−0=30). PASS98. Undercover rule: 0.844 < 0.95−3·0.004873=0.93538 -> true; also robust to using SE at p̂ (0.95−0.844=0.106 ≫ 3·0.0081). PASS109. Magnitude vs baseline: mean se_conv (0.190) < exact SD (0.268), ratio 0.71 — explains undercoverage; HC1 mean SE 0.241 closer but still below → HC1 also undercovers (0.902). Coherent. PASS11