# Suite23 frozen design rationale

## Evidence entering Suite23

Suite20–21 showed that historical academic content is useful, but precise semester ordering becomes a modest contributor once a strong tabular anchor exists. Suite22 then supplied a much stronger result: the student-course relational branch was the only major architecture component with a large positive increment, while faculty modulation, temporal feature modulation and the sparse scalar feature graph did not generalize robustly.

Suite23 therefore gives the neural model a narrower responsibility: **learn relational academic evidence that the Committee cannot represent directly.**

## Relation to DGTEAM

The attached DGTEAM paper motivates three dynamic structures: student-course interactions, student similarity and course similarity. Suite23 keeps the central insight but redesigns the implementation for AQ:

- student similarity is induced through Student→Course→Student paths and the target-course cohort readout, not a dense handcrafted JSD student graph;
- course similarity is induced through Course→Student→Course paths plus explicit prerequisite/related-course edges;
- raw IDs retrieve neighborhoods but are not learned identity embeddings;
- semester sequence modeling is replaced by relative-time edge attributes;
- continuous-grade regression is replaced by failure classification;
- curriculum/course-structure relations are explicit typed edges;
- Committee-preserving fusion is predeclared as the deployment path.

## Fixed query graph schema

Node slots per query: 31.

- 1 root student
- 1 root target course
- 1 faculty
- 8 historical courses
- 12 target-course historical students
- 6 structural/prerequisite courses
- 2 structural course-group nodes

Maximum edge slots: 68.

## Why two layers

Two layers permit the central latent paths needed by the hypothesis without deeper oversmoothing:

- Student → historical Course → Target Course
- Course → historical Student → implicit student-similarity evidence
- Student/Course → Faculty → cross-context interaction

The dual readouts then directly query the two relational evidence sets with the opposite root representation.

## Falsification criteria

The graph hypothesis is weakened if all of the following occur:

1. `heterogeneous_graph - bipartite_graph ≈ 0`;
2. `cs_hagt - heterogeneous_graph ≈ 0`;
3. zero/shuffle relation perturbations have little effect;
4. graph support does not correlate with graph gain;
5. Committee+CS-HAGT does not exceed Suite22 safe blend.

The implementation deliberately prints each of these signals so the experiment can fail informatively.


## Implementation safeguards added during finalization

- Course-relation direction is preserved with the same forward/reverse typed coding principle used by the existing structured cache; exact same-course history is explicit rather than inferred from similar features.
- Faculty graph nodes use a learned shared faculty embedding rather than treating the integer faculty index as a continuous numeric feature.
- Structural-group nodes aggregate member-course profiles and do not rely only on group size/ID.
- History-shuffle perturbations preserve target course and faculty whenever possible, and replace the complete history subgraph evidence rather than only changing feature values.
- Production refit reuses the selection seed to reduce gratuitous calibration shift between chronological selection and full-history refit.
- Non-monotone Platt calibration fails safely to identity rather than inverting the graph ranking.
