# Verification checks (P4)
1. Clean-state rerun of out/analysis.py in /tmp copy -> results.json byte-identical. PASS
2. Counts: n_analysed 1200 = n_firms 100 x n_years 12 = n_input_rows 1200 - 0 removed (exclusions.csv). PASS
3. se_conventional > 0, se_cluster > 0. PASS
4. CI [0.5151, 0.6938] contains beta_fe 0.6045 and is symmetric (beta_fe +/- 1.984*se_cluster). PASS
5. FWL identity: beta from residualised x equals LSDV beta (assert in script, tol 1e-8). PASS
6. Cluster SE cross-check: our all-K CR1 SE equals statsmodels LSDV cluster SE (0.047057 both). PASS
7. Magnitude vs baseline: beta_fe (0.60) differs from beta_pooled (1.27) -> consistent with FE correlated with x; plausible. se_cluster/se_conventional = 1.38, consistent with within-firm serial dependence. PASS