# Suite12 — Collaborative Evidence MoE with Distilled Shared Expert

## Research question
Can a fully differentiable, direct-prediction collaborative MoE substantially outperform the frozen tree Committee and Suite4 Competence Council by combining genuinely heterogeneous academic evidence rather than residual corrections?

## Core design decision
The RF/ExtraTrees/HGB Committee is **not Expert 0 in the deployed graph**. It is a frozen teacher and evaluation baseline. E0 is a differentiable neural global expert pretrained to absorb the three individual teacher views, then jointly fine-tuned with the collaborative system.

## Evidence experts
- **E0 global** — full common flat representation + broad chronology-safe support context. Multi-head teacher distillation + direct failure head.
- **E1 trajectory** — semester sequence from `DDTBHK`, GRU + temporal attention. Student-warning/conduct/workload/history context only.
- **E2 curriculum** — target course, chronology-safe population course history, and one-hop prerequisite/curriculum graph. Local relation-aware neighbor attention. No target-student faculty embedding or student-centric context.
- **E3 compatibility** — chronology-safe historical student-course outcomes, prioritized by same course → direct prerequisite → related → other recent courses, with explicit relation-direction/type encoding.
- **E4 mastery** — chronology-safe historical component events from `DDiemTPTG`, component-type embedding, recency, same-course/curriculum relation, attentive aggregation.

## Context isolation
Private experts do not receive E0 or Committee representations before private reasoning.

The final fusion/gate receives only neutral evidence-support context:
`history_semesters`, `history_rows`, `course_prior_rows`, `new_student`, `new_course`.
This prevents warning/conduct/workload variables from bypassing the experts through the final head.

## Collaboration
Five private tokens are passed to a small gated evidence-exchange layer. Each expert retains private evidence, received messages, and disagreement (`private - message`). Communication strength starts small. Fusion is dense rather than Top-K because there are only five experts and previous suites showed brittle row-ownership routing.

## Prediction
The final output is a direct failure probability. There is no Committee residual anchor and no exact fallback mixed into the candidate score space.

## Data / chronology
Reuse the Suite5 canonical cache for flat X, labels, OOF teacher predictions, blocks, support, and evaluation alignment. Build a separate structured raw-evidence index/cache for E1–E4. Dynamic events must satisfy `NHHK_event < NHHK_target`. Timestamped curriculum edges must also predate the target semester.

## Training
0. Distill/train E0.
1. Pretrain E1–E4 privately.
2. Train evidence exchange/fusion with private encoders frozen.
3. Jointly fine-tune every neural module end-to-end with weak continuing E0 distillation.

Structural evidence separation is the primary anti-collapse mechanism. Additional protection: private auxiliary heads, expert dropout including E0 dropout, small initial communication strength, and comprehensive diagnostics rather than strong decorrelation penalties.

## Success criterion
Suite12 is successful only if it produces substantial forward gains across a strong majority of faculties, including AP and operational F1/F2/confusion-matrix improvements. For IT, Suite4 Council remains the reference to beat. Architectural diversity, gate usage, or low representation similarity are diagnostics—not success by themselves.


## Data provenance clarification (v4)

Suite12 treats the engineered expanding folds / canonical Suite5 cache as authoritative for row identity, labels, flat E0 input, temporal blocks, and fair evaluation. The raw database does not redefine the dataset. It contributes only strictly historical structured evidence for E1-E4 after a mandatory ID/semester/grade alignment audit. Extracted database-directory input is preferred; ZIP input is a portability fallback. DDTBHK duplicate `(IDSV,NHHK)` summaries are resolved by latest `NgayThem`, which reproduces the engineered fold's previous-semester GPA semantics in the alignment audit.
