# Suite13 Committee-Preserving Collaborative Evidence MoE — Design

## Research question
Can Suite12's genuinely heterogeneous, densely collaborative evidence architecture surpass the tree Committee and Competence Council when the global expert preserves the Committee nearly losslessly, model selection is cross-faculty/temporal-robust, joint fine-tuning cannot erase E0, specialists receive mechanism-specific auxiliary supervision, and the gate can condition on known faculty/semester regime information?

## Fixed architecture from Suite12
- E0 global neural/tabular evidence expert.
- E1 student semester-trajectory expert.
- E2 course/curriculum expert.
- E3 student-course compatibility expert.
- E4 fine-grained component/mastery expert.
- Private reasoning occurs before expert exchange.
- Dense gated expert-to-expert representation exchange.
- Dense final fusion; no Top-K row ownership.
- Final prediction is direct, not Committee + MoE residual.
- Existing Suite12 structured cache schema remains reusable.

## Modification 1 — Committee-preserving E0
The actual chronology-safe RF/ET/HGB probabilities are direct frozen inputs to E0. E0 forms a teacher-evidence vector from individual probabilities/logits, Committee mean, disagreement spread/std and Committee entropy. A neural flat-feature branch and teacher-evidence branch are fused into h0.

E0's private risk logit is:

    z0 = logit(q_committee) + delta0(h0)

where delta0 is zero-initialized and bounded. Thus E0 equals the Committee before optimization but can learn a direct global correction. This residual is internal to E0 only; Suite13 final output remains direct collaborative fusion.

Phase 0 and Phase 3 use explicit Committee-preservation loss in addition to label supervision. Teacher reconstruction remains auxiliary so h0 retains the individual RF/ET/HGB views.

## Modification 2 — Cross-faculty/temporal robust selection
Phase 0, 2 and 3 epochs are selected on the chronology-safe reliability block using faculty-level deltas versus the Committee, not pooled AP. Primary robust score:

    median(delta AP) + w_f2 * median(delta F2@20) + w_mean * mean(delta AP)

Diagnostics include AP/F2 win counts, median/mean deltas, worst-quartile AP delta and E0 retention. Phase 1 uses faculty-balanced mechanism/private validation loss rather than pooled classification AP.

The newest historical selection block remains a separate untouched audit and never changes target predictions or creates fallback score mixing.

## Modification 3 — Protect E0 during joint fine-tuning
Phase 3 uses optimizer parameter groups: E0 learning rate is joint_lr * e0_lr_scale while all other trainable modules use joint_lr. Committee-preservation and teacher-reconstruction losses remain active. Diagnostics record E0-vs-Committee AP, MAE, rank correlation, top-20 overlap and residual-logit magnitude after every selected/refit model.

## Modification 4 — Mechanism-specific auxiliary objectives
Private failure heads remain for all experts for comparability, but E1-E4 additionally receive mechanism supervision:
- E1: chronology-safe recency-weighted student academic-risk summary computed only from prior-semester GPA/credit/debt evidence.
- E2: chronology-safe historical course failure propensity already present in the course-history tensor; unseen courses receive a neutral target.
- E3: row-level student-course failure target.
- E4: chronology-safe mastery-deficiency target computed from historical component events, weighting low/passed status, recency and target-course/curriculum relevance.

E1/E2/E4 auxiliary targets are evidence-derived rather than same-semester label aggregates, so no peer label from a reliability/selection block can enter a training auxiliary target.

Direct per-row failure supervision is retained with expert-specific weights so E2 is not punished for being intentionally course-centric.

## Modification 5 — Domain/time-aware gate
The gate, but not the final risk head, receives:
- faculty embedding,
- standardized target-semester ordinal,
- evidence-support variables (history depth, course support, new student/course flags).

This allows expert trust to vary by faculty and temporal regime without introducing an independent sixth risk predictor. Final risk head still receives only fused evidence plus neutral support context.

## Evaluation and diagnostics
Use the same evaluator for Committee, V2, Council, Suites7-12, Suite13 E0, E1-E4 and Suite13 final. Report AP/AUC; P/R/F1/F2 at 10/15/20/25%; fixed and drift-aware policies; TP/FN/FP/TN; faculty win rates.

Additional Suite13 diagnostics:
- historical robust-selection score components per epoch;
- E0 Committee retention: AP delta, probability MAE, rank correlation, top20 overlap, logit-correction mean/p95;
- RF/ET/HGB reconstruction MAE;
- private failure AP and mechanism-target BCE/MAE/correlation;
- collaboration delta vs E0 and best private expert;
- gate entropy/utilization overall/failure/normal and gate temporal/faculty context;
- expert ablation delta AP/F2;
- private/collaborative representation cosine;
- all/failure/normal attention matrices;
- top20 rescued/lost failures and false-alarm transitions;
- comparison and win-rate audit vs Suite12 and earlier systems.

## Success criterion
Mechanistic improvements alone are not success. Suite13 must show broad cross-faculty forward AP/F1/F2 improvement and materially challenge or exceed the IT Council (mean-fold AP 0.4449, drift F2 0.5477) without exploding false alarms.
