Verification checks (P4), 2026-10-29
- Clean re-run (deleted results.json, exclusions.csv; re-ran analysis.py): byte-identical results.json — PASS
- CI contains beta_fe: PASS
- SEs positive: PASS
- n_analysed = n_firms*n_years = 1200 (balanced): PASS
- n_input_rows (1200) - total removed (0) = n_analysed (1200): PASS
- Two-way within-demeaning beta equals LSDV beta (|diff|<1e-8, also asserted in script): PASS
- CI half-width = t(0.975, G-1=99) * se_cluster: PASS
- LSDV parameter count 112 = const + 99 firm + 11 year + x: PASS
- Magnitude vs baseline: beta_fe (0.604) is about half of beta_pooled (1.268): consistent with firm/year effects correlated with x; not an error signal. se_cluster/se_conventional = 1.44, plausible with within-firm serial correlation.