# Suite32 validation record

**Environment:** local Python/scikit-learn integration using generated synthetic AQ-shaped joblib caches only; installed packages in the current container. This is **not** a university-data execution.

Tests in `tests/test_suite32.py` cover:

- Valid semester ordinals and dynamic common/native schema alignment.
- Donor-native features absent as `NaN` rather than invented numeric values.
- Cache NaN indicator different from naive `value==0` indicator.
- Global chronology violation raises an error before training.
- Canonical cross-fitted historical row index mismatch raises an error.
- Weighted target/donor loss mass normalizes; positive and total ESS are logged.
- Historical source-selection candidate screening excludes one-origin apparent gains.
- Prespecified lexical language-faculty grouping control.
- Binary classification AP and paired faculty-fold comparisons.

**Integration:** generated 3-faculty caches (8 shared raw features + 17 structured course features + 3 faculty-native features). Ran the complete `train_suite32.py` command from loader through source screening, nested replay, candidate admission, controls, `--mask-audit`, row-level predictions, reports and diagnostics. The source-selection fallback triggered on a negative mean replay uplift as expected. The synthetic AP values are *not* comparable to the 14-faculty university experiment.

**Before publishing real AP:** run `preflight` and full F1–F5 development, audit feature names/dimensions, inspect donor decision logs, freeze `rho`, support sizes and hyperparameters, then F6/F7 retrospective. Verify original grading event availability and raw source missingness independently if claiming prospective or provenance-safe results. Align archived Council/Suite25 row-keyed predictions before student-level paired CIs. Obtain genuinely later semesters for untouched prospective tests.

## Logging/output contract verification (2026-10-17)

- `run_suite32.sh preflight --folds 1 --max-targets 1` against synthetic caches:
  terminal and stage log matched byte-for-byte; canonical log and uniquely
  timestamped per-invocation log created under `<AQ_ROOT>/log/`.
- Intentional invalid `--rho 1`: Python traceback and `[FATAL]` captured in
  all logs; runner exited non-zero with `[RUN STATUS] FAILURE`.
- Synthetic `development --folds 1 --max-targets 1 --max-iter 5`:
  stage log exactly matched terminal; real per-faculty, transfer, replay,
  in-sample training, source mass, macro, report and insight printings were
  captured. Predictions, model diagnostics and report CSVs were written to
  `<AQ_ROOT>/results/suite32_selective_borrowing/`.
- `python -m unittest discover -s tests -v`: **10 tests passed**, including
  2 regression tests for runner logging.
- No real university data were used for this verification.
