# Suite14 — Authority-Safe Collaborative Evidence MoE

Suite14 is the final conservative continuation of the Suite12/13 heterogeneous collaborative framework.
It reuses the existing canonical cross-faculty cache and the Suite12 schema-v4 structured cache.

## Research question

Can E1–E4 add complementary structured evidence **without being allowed to damage a strong E0 by default**?

Suite14 keeps:
- E0: Committee-preserving global expert using RF/ET/HGB evidence.
- E1: student trajectory GRU/attention.
- E2: course/curriculum expert.
- E3: student-course compatibility expert.
- E4: component/mastery expert.
- gated expert exchange and mechanism supervision.

Suite14 changes:
1. Final prediction is identity-preserving: `z = z_E0 + authority * row_alpha * delta_collab`.
2. Collaborative correction is zero-initialized; Phase2 epoch 0 is exact E0.
3. E0 is frozen after Phase0 and cannot be damaged by later joint collaboration training.
4. The gate does **not** receive absolute target-semester ordinal. It uses neutral support, faculty, expert probabilities, disagreement, confidence, and availability.
5. Phase1 checkpoint selection combines mechanism fidelity with actual private failure-ranking AP.
6. Authority is calibrated on older chronology-safe validation blocks and confirmed on the newest held-out historical block.
7. Authority 0 is always available. If broad validation support is absent, only faculties with strong local temporal evidence may retain nonzero authority; all others fall back exactly to E0.

## Output score variants

Every faculty score file includes:
- `suite14_e0_global`: protected E0.
- `suite14_full_authority_candidate`: raw model with external authority = 1.
- `suite14_calibrated_candidate`: per-faculty authority selected on historical calibration blocks.
- `suite14_authority_safe_final`: operational deployment after newest-block confirmation/broad-success policy.
- `suite14_private_e1` ... `suite14_private_e4`.

This distinction is intentional: it lets us determine whether the collaborative hypothesis is useful but over-authorized, or fundamentally weak.

## Authority diagnostics

Each fold writes:
- `authority_calibration.csv`: selected authority, temporal calibration gains, newest-block confirmation gains, deployment authority, fallback reason.
- `historical_selection_audit.csv`: E0, full collaboration and final authority-safe gains on held-out historical blocks.
- `suite14_diagnostics.csv`: E0 retention, full/calibrated/final AP, row-authority distribution, correction magnitude, expert gate usage, ablations, mechanism diagnostics, failure rescue/loss transitions.
- attention/representation diagnostics and the same-evaluator operational reports.

## Run

Expected layout:
- canonical cache: `/root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe/cache`
- structured cache: `/root/disk_3tb/earlywarning/30072026/results/suite12_structured_cache`
- raw DB fallback: `/root/disk_3tb/earlywarning/database`

```bash
cd /root/disk_3tb/earlywarning/30072026/suite14
bash run_suite14.sh
```

Default results:

```text
/root/disk_3tb/earlywarning/30072026/results/suite14_authority_safe_collaborative_moe
```

Default log:

```text
/root/disk_3tb/earlywarning/30072026/log/suite14.log
```

No Suite12 structured-cache rebuild is required if schema-v4 files already exist.
