# Suite13 — Committee-Preserving Collaborative Evidence MoE

Suite13 keeps Suite12's heterogeneous E1–E4 evidence encoders and dense expert collaboration, but implements the five changes motivated by the Suite12 postmortem.

## What changed from Suite12

1. **Committee-preserving E0.** The actual chronology-safe RF/ET/HGB probabilities are direct frozen evidence. E0 starts exactly at the Committee probability through a zero-initialized bounded internal correction. The final Suite13 prediction is still direct collaborative fusion, not Committee + final residual.
2. **Robust temporal/cross-faculty selection.** Phase 0/2/3 epochs are selected by faculty-level Committee-relative median ΔAP + weighted median ΔF2@20 + weighted mean ΔAP. Pooled AP no longer chooses epochs.
3. **Protected E0 joint tuning.** Phase 3 uses a lower LR for E0 and keeps Committee-preservation/reconstruction losses active. Phase 3 can retain epoch 0 if joint tuning hurts the robust validation score.
4. **Mechanism-specific auxiliary objectives.** E1 learns a prior-semester academic-risk summary, E2 historical course fail propensity, E3 row failure, E4 component mastery deficiency. E1/E2/E4 targets are chronology-safe evidence-derived summaries, not peer-label aggregates.
5. **Faculty/time-aware gate.** The gate sees faculty plus standardized target-semester ordinal and evidence support. The final risk head does not directly see faculty/semester; those variables affect expert trust only.

## Data/cache policy

Canonical rows, labels and RF/ET/HGB OOF/target probabilities still come from:

`/root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe/cache`

Suite13 reuses Suite12's structured cache by default:

`/root/disk_3tb/earlywarning/30072026/results/suite12_structured_cache`

No raw database scan is needed when that compatible schema-v4 cache already exists. If it is missing/stale, `run_suite13.sh` can rebuild it from `/root/disk_3tb/earlywarning/database`.

## Run

```bash
cd /root/disk_3tb/earlywarning/30072026/suite13
bash run_suite13.sh
```

Log:

`/root/disk_3tb/earlywarning/30072026/log/suite13.log`

Results:

`/root/disk_3tb/earlywarning/30072026/results/suite13_committee_preserving_collaborative_moe`

## Key diagnostics

Each fold prints/saves:
- robust per-phase Committee-relative selection metrics;
- newest historical selection audit;
- E0 Committee retention: ΔAP, probability MAE, rank correlation, top-20 overlap, logit correction magnitude;
- private E1–E4 AP;
- E1–E4 mechanism BCE/MAE/correlation;
- collaboration gain vs E0 and best private expert;
- gate entropy/utilization overall and by outcome;
- expert ablation ΔAP/ΔF2@20;
- representation similarities and private→collaborative shifts;
- all/failure/normal attention matrices;
- top-20 rescued/lost failures and false-alarm transitions;
- same-evaluator comparisons against Committee, V2, Council, Suites7–12.
