# Data audit (P1)

Hashes match bundle.json.

Schema: cell_id (str C1..C9), G (int 5/10/30), icc (float 0.0/0.2/0.5), cluster_sizes (str, '|'-separated ints), N (int), R (int, 2000 all rows), B (int, 399 all rows), seed (int 20261101..20261109), mu (0.0), total_variance (1.0), nominal_level (0.95). Units: dimensionless (y has unit variance).

Checks: no missing values; no duplicate rows or cell_ids; for every row len(cluster_sizes)==G and sum(cluster_sizes)==N (128, 264, 864). Sum n_g^2 = 5504 / 11200 / 38400. Grid is the full 3x3 factorial. No dates (simulation design), so no temporal availability issues; no input violates the information rules.

Anomalies / design notes (not errors):

• ICC=0.0 cells: u_g ~ N(0,0) is degenerate (u_g = 0); implemented as 0*standard normal draw so the RNG stream structure is identical across ICC.
• G=5: Rademacher weights have only 2^5 = 32 distinct vectors, and |t*| is invariant to w -> -w, so at most 16 distinct |t*| values; the WCB critical value is coarse. Kept as specified (no change to procedure), reported in data_issues.
• The cluster-size patterns are unbalanced (G=5 repeats size 8; G=10 and 30 are truncated repetitions), as documented in README.

Exclusions: none (all 9 cells simulated). See out/exclusions.csv.