# Suite30 v1.0 — Release checklist

## Included / verified
- [x] Four jointly comparable training configurations (independent fits).
- [x] Shared semester latent-risk mixture and exact count distribution.
- [x] Leave-one-out peer attention; coherent marginal and joint outputs.
- [x] BCE + NLL + any/multiple/count + ranking auxiliary losses, ramp and recency.
- [x] Faculty-balanced sampling, global chronological validation, full-history refit.
- [x] Historical-only, faculty-shrunk intercept calibration preserving joint PMF.
- [x] No-attention and no-context interventions; exact-match peer-course placebo.
- [x] Matched-only placebo AP effect, faculty and set-size error attribution.
- [x] Exact-key verified Suite29 comparison, Committee baseline.
- [x] Cross-fold AP/F2/Brier, group counts, any/multiple, joint NLL diagnostics.
- [x] Checkpoint with preprocessing/calibration; replay all interventions.
- [x] One stable logfile in project `log/` directory; overwrite per execution.
- [x] 19 model/data/chrono/calibration unit tests.
- [x] Synthetic canonical/structured joblib training tested (2 faculties, 2 folds).
- [x] Synthetic checkpoint exact-row replay passed.

## Not yet established
- [ ] Execution against actual 14-faculty university cache files on the user's server.
- [ ] Upstream feature as-of timing provenance.
- [ ] Empirical AP improvements versus Suite29/Committee on real F1–F7.
- [ ] Future untouched validation beyond previously consulted F6/F7.

Read `README.md` and `DESIGN_AUDIT.md` before claiming scientific results.
