# Suite33 v2 Release Notes

## Added

- New `--mode reconstructed` default: database-backed exploratory course/section context from four existing CSVs; no fictitious roster manifest.
- `--database-dir` option (defaults to `<root>/../database`).
- Source provenance / missingness / conflicting-registration diagnostics, per-origin quality rates, SHA-256 data hashes and cautionary machine-readable scientific status.
- Explicit non-deployment guard even if S4 outperforms S1 in development.
- Retrospective target-subset sensitivity using records present by the semester-start proxy. The underlying peer contexts remain reconstructed; this is NOT historical-roster certification.
- Four reconstruction-specific automated tests + synthetic raw database generator.
- S4 course-wide relative percentile included as originally documented (P1 holds it fixed).

## Preserved

- S0–S4 and P1; exact canonical covered-row/sidecar alignment; fixed HGB; 14 faculties; all F1–F7; fold/faculty AP/F2/AUC/calibration and detailed diagnostics; one launcher and one log; certified mode optional.

## Explicit limitations

- Reconstructed CSV source is current state and may have future enrollment membership. Source `MKQDK.IDDVPC` is a unit code, NOT a verified faculty label.
- `DDTBHK.NgayThem` is a conservative insertion-time proxy, not proof of grade publication.
- The previous admission criteria are development research signals only in reconstructed mode; `promoted_for_deployment=false` regardless of AP.
- No real dataset model results were computed in this deliverable. End-to-end validation uses synthetic AQ-like records.
