# Suite22 frozen design — SFTR-TabM

## Primary hypothesis

A modern parameter-efficient shared tabular learner can outperform the previous specialist/residual line when it receives four forms of structured inductive bias: nonlinear feature interactions, faculty-dependent feature semantics, continuous temporal concept-drift context, and student-course relational evidence.

## Primary architecture

`current scalar features -> training-only piecewise-linear tokenizer -> low-rank faculty/time/faculty×time modulation -> sparse feature graph -> concatenate relational ego-graph state + faculty/time context -> official TabM K=32 -> ensemble mean probability`

### Current-state representation

- common raw features shared by all 14 faculties;
- 17 target-course static/history statistics;
- each scalar becomes a 4-D piecewise-linear token using training-only quantile hinge knots.

### Faculty/time modulation

The modulator is decomposed as `M_F + M_T + M_FT`. A low-rank coefficient system produces bounded per-feature scale, shift, and nonlinear-shape corrections. It initializes to exact identity.

### Time representation

No semester lookup table. Use normalized ordinal trend plus sin/cos periods 3, 6, 12. This is intended to model evolving feature semantics and remain defined on future folds.

### Feature graph

A global learned feature-identity graph keeps the top 8 outgoing relations per feature. Message passing has a small learned authority initialized near 0.1. This is T2G-inspired, not an implementation of T2G-Former.

### Academic ego graph

The target course queries four structured evidence sets: prior semester states, related/prerequisite course nodes, prior relation events, and component/mastery events. These sets are permutation invariant apart from recency already encoded inside the structured evidence.

### Final TabM

Production uses official `tabm==0.0.3`, `K=32`, 3 blocks, width 512, dropout 0.1. The external frontend is trained end-to-end through TabM.

## Controlled variants

A Shared TabM -> B Faculty Mod -> C Faculty-Time Mod -> D Relational FT-TabM -> E Full SFTR-TabM.

The primary result is E. Holdout outcomes may not be used to retrospectively promote another variant.

## Optimization

- AdamW lr=0.002, wd=3e-4
- batch 512
- max 15 epochs, min 3, patience 3
- faculty-balanced sampler only
- BCE across TabM members + 0.10 within-faculty pair ranking regularizer
- chronological latest-3-block validation
- epoch selection lexicographically by macro AP, F2@20, then Brier

## Holdout diagnostics

The primary model receives post-hoc `no_faculty_mod`, `no_time_mod`, `no_interaction_mod`, `identity_mod`, `graph_off`, `zero_relations`, and within-faculty `shuffle_relations` interventions. It is also tested at counterfactual semester `t-1` and `t+1` for stability. None of these affect model selection.

## Evaluation

All earlier available models are loaded row-aligned where possible. F1–F5 define frozen operational thresholds and warning-rate ratios. F6/F7 report ranking, calibration, budget metrics, confusion matrices, and 200-replicate paired bootstrap ΔAP against Committee, Suite18 and Suite21 Guarded.
