Data audit (P1)
| file | SHA-256 | rows | cols |
|---|---|---|---|
| data/design.csv | f4ff004ed50d893eece1aaa7edced04d3c0439aaedf1d391ea4642a87d0ecc50 | 9 | 11 |
| data/README.md | 73a6595909d92a69db281e283f193348ac244f255fcd5343ef42ae0353394e67 | (text) | — |
Hashes match bundle.json. design.csv schema: cell_id (str), G (int), icc (float), cluster_sizes (str, '|'-separated ints), N (int), R (int), B (int), seed (int), mu (float), total_variance (float), nominal_level (float).
- Missing values: 0. Duplicate rows: 0. Seeds unique (20261101–20261109).
- For every cell, number of listed sizes = G and sum of sizes = N (128, 264, 864). Σn_g² = 5504, 11200, 38400.
- Units: y dimensionless, total variance 1. No dates (synthetic design; no chronology).
- Anomalies / notes: (i) ICC=0 cells give u_g with variance 0 (degenerate, handled by numpy normal with scale 0 → exactly 0). (ii) Cluster sizes are unbalanced (8–64); with G=5 the 64-cluster holds 50% of observations — high leverage, expected to hurt sandwich estimators. (iii) With G=5 Rademacher weights generate only 2^5=32 distinct weight vectors (16 distinct |t*| values) so the WCB critical value is coarse; this is a property of the specified design, not an error.
- Information-availability: no input violates the rules (design only, no observed data). No exclusions are needed.