# Suite33 v2 Research Protocol — Frozen exploratory reconstruction study

**Updated after the October 18, 2026 database-source timing audit.** This document supersedes the origin-provenance assumptions in the original `RESEARCH_DESIGN.md`, which remains archived for history and completeness.

## 1. Scientific question

Can prior-semester academic preparation of *other students in a target course/section*, assembled from the latest available institutional raw export, improve per-registration course-failure ranking beyond the existing 153-feature AQ benchmark?

This is an exploratory **associational information study**, because the latest available `MKQDK`/`KToHoc` exports do not record an immutable semester-origin roster. It is **not** the stronger claim that the same context could have been obtained at the official semester prediction time.

The binary outcome remains `numeric grade <5` for the original canonical eligible registration population. The pre-existing fold definitions, student/course keys, 14 faculty labels, F1–F5 development and reused F6/F7 retrospective evaluation remain fixed. Changing to a later prediction horizon, using current-term grades, or filtering out original benchmark targets would change the estimand and is not allowed.

## 2. Direct source evidence and its limits

October 2026 audits established:

- 1,378,743 `MKQDK` rows, no duplicated `IDKQDK`, complete `IDSV`, `IDToHoc`, `NHHK`, `NgayThem`.
- 33,037 `KToHoc` rows, no duplicate `(NHHK,IDToHoc)` mapping, almost perfect course/section match.
- `HDMNHHK.NgayBDHK` is populated, but cannot be equated to complete roster availability.
- Among the audited 20221–20252 semesters, approximately 77% of registrations and 65% of sections were recorded after their configured term-start date; 63% of recorded section changes were after the cutoff. Many individual semesters exceeded 99% late creation.
- `DDTBHK.NgayThem` is populated but is not independently verified grade publication. Some earlier-semester records are inserted after later-term cutoff.

The current source therefore supports a reconstructed-roster *sensitivity experiment*, not a proof of historical prediction availability.

## 3. Reconstructed dataset and exclusions

Join `MKQDK(NHHK,IDSV,IDToHoc)` to `KToHoc(NHHK,IDToHoc,IDMH)` and `HDMNHHK(NHHK,NgayBDHK)`. Normalize keys as strings; interpret local naive timestamps as `Asia/Ho_Chi_Minh`. Exclude records with no section mapping, missing cutoff, or invalid required keys; count and report every exclusion. If one student has more than one exported section for the same course in a semester, exclude **all** rows with that ambiguity and report counts; never arbitrarily choose a section.

Use the complete exported registration population to represent current-state classmates, even if their outcome is not eligible for the target benchmark. Retain the original benchmark target rows independently, including when no roster membership is matched (extra feature block becomes missing). Record membership availability by faculty and fold.

Join `DDTBHK` by student, using a prior-semester GPA with semester earlier than target and `NgayThem` strictly earlier than the official semester-start proxy. Select latest admissible previous semester, and then latest entry within that semester; group-history GPA coverage remains observable. Do not equate `NgayThem` to certified publication.

The aggregate uses **no target-semester outcomes** and excludes the target student from course-wide and section-level peer statistics. No model-input feature can depend on whether a registration has an insertion date after origin, a later section-change date, or other current-state provenance flags.

`MKQDK.IDDVPC` stratifies placebo by a source administrative unit. This is not a verified student faculty label and must not be interpreted as one.

## 4. Predeclared arms

S0 = matched canonical own-history baseline.

S1 = S0 + section and course size, previous-GPA coverage/counts and source-group diversity.

S2 = S1 + leave-self-out course-wide prior-GPA mean and standard deviation.

S3 = S2 + leave-self-out reconstructed section prior-GPA mean/SD and the contrasts versus course-wide context.

S4 = S3 + target student's own previous-GPA gap/normalized gap and relative section/course percentiles.

P1 = S4 with a deterministic, within `(semester,course,administrative-unit)` reassignment of section identifiers that preserves section size within each source unit. Course-wide context remains unchanged. Only groups containing multiple real sections permit an informative section-placebo interpretation. All arms use an identical fixed HGB learner with 100 iterations for real 14-faculty comparisons.

Missing peer GPA is missing, not GPA zero. Missing derived fields receive the same numeric sentinel policy across all arms. Context-derived features never use current-semester target grade labels.

## 5. Prespecified comparisons

Primary descriptive question: S4 minus S0, macro AP averaged equally across faculty-fold cells. Mechanism increments: S2−S1 (course composition), S3−S2 (section incremental composition), S4−S3 (relative position), S4−P1 on matched multi-section subsets (actual section vs shuffled membership). Use full faculty/fold matrices, cell win/loss counts and worst harms; no selective spotlight only on IT.

Report AUC, F1/F2, per-faculty historical thresholds, alert rate, calibration and uncertainty sensitivity. Baseline comparisons to Suite25/Council archives are descriptive cross-run metrics until actual row-paired scores are available. An S4 selection based on F1–F5 is a **research-level conditional signal only**; it can never be promoted as validated origin-time deployment based solely on reconstructed sources. F6/F7 were inspected in earlier suites and are not untouched validation.

## 6. Reconstruction uncertainty diagnostics

Emit per semester and per target faculty/fold:

- Raw registration count; unresolved joins, missing origins, ambiguous student/course entries.
- Fraction inserted after semester-start proxy.
- Fraction of sections created after that proxy.
- Fraction with recorded later section changes; inability to reconstruct prior section assignments.
- Fraction of peer histories with usable earlier GPA based on insertion proxy.
- Target registration roster match rate; section support and source-unit diversity.
- Full S0–S4 and P1 predictions; output provenance warnings; source content SHA-256 hashes.
- One narrowly qualified diagnostic subset of target rows with records existing by origin. **All context features for this subset remain constructed from the full current-state roster**, so subset results do not prove historical validity.

If historical rosters become independently available, rerun the original strict `--mode certified` arm as a separate, clearly identified protocol; never upgrade exploratory manifests by asserting unverified booleans.

## 7. Execution and stop conditions

One `./run_suite33.sh`, one appended log, chronological F1–F5 followed by F6/F7, detailed per-cell training and research insights. If the required raw tables or canonical caches/sidecars are missing, or key/feature alignment fails, abort with the cause rather than silently substituting the final eligible-grade fold roster.

If the exploratory context yields no consistent increment over S0/S1 or matched placebo, close this direction. If it yields positive evidence, prioritize acquiring historical enrollment snapshots and verified grade publication records as the **next data-collection step**, not claiming deployment based on the retrospective reconstruction.

## 8. What is not implemented or claimed

No full 10,802-row recovery experiment, no raw-split missingness 2×2, no validated student faculty crosswalk, no reproducible instructor-version join, no causal peer-effect identification, and no real AQ performance results are included in the deliverable. They are separate follow-up questions and should not be inferred from program completeness.
