# Historical AQ result provenance

These two CSVs are copied unchanged from the supplied `suite32_selective_borrowing(1).zip`, directory `suite32_selective_borrowing/reference`. The supplied Suite33 package contains the same reference filenames. They are archived measurements, not regenerated Suite35 evidence.

- `historical_model_ap_cells.csv`:2,261 faculty/fold/model cells,14 faculties, folds1–7,24 model labels. It includes Committee, Suite22, Suite25/blend and other earlier variants. The `evidence` column records the original provenance. Available target row/failure counts are retained.
- `council_13_faculties.csv`:91 cells across13 faculties and7 folds, with archived council AP/AUC and warning-policy metrics. It is retained as supporting context; it is not merged as if it covered all14 faculties.

The default Suite35 report matches the first file by faculty/fold, averages current seed AP within the cell, then averages matched cells equally. It reports estimable/matched cell counts, AP deltas, wins and target-population count agreement. Matching counts do not verify identical row identities, identical features or identical training support. These files contain no aligned probabilities, so comparisons are descriptive, cannot enter row-paired bootstrap contrasts, and never determine Suite35 promotion. A partial set of faculties is never represented as the complete14-faculty baseline.

No historical scores were invented or estimated from plots. The original reference data remain in these CSVs; missing cells remain missing.
