# Suite10 — Semantic Academic-Mechanism Specialists

## Research question
Can observable, pre-outcome academic mechanisms define stable expert responsibilities that preserve Suite9's functional specialization while making routing interpretable and temporally learnable?

## Fixed backbone
Reuse Suite7's cross-faculty shared encoder, faculty-specific bottleneck adapters, frozen Committee anchor, four residual expert slots, chronological OOF cache, and evaluation protocol. Suite10 initializes from Suite7; only residual experts are specialized. No neural router is trained in the first semantic-domain experiment.

## Semantic expert domains
- E1 Student-sparse/cold-start: weak or absent student academic history.
- E2 Course-sparse/curriculum-novel: weak or absent prior course-side support.
- E3 Temporal deterioration: recent academic state is worse than longer-term state.
- E4 History-rich relational: substantial student history and enough course/relational evidence to support student-course-specific reasoning.
- E0 NULL/Committee: no validated specialist mechanism warrants authority.

All domain membership scores are computed exclusively from pre-outcome features and chronology-safe support. Labels never decide domain ownership.

## Soft domain responsibility
Each domain produces m_k(x) in [0,1]. Experts train with domain-weighted counterfactual BCE, normalized separately per expert. Soft overlap is allowed; a row can train multiple mechanism experts. Core-domain analysis uses m_k >= 0.60 by default.

## Temporal deterioration score
Use only observable recent-vs-long-term deterioration signals: semester GPA, course grade, component grade, latest-vs-cumulative GPA, warning-rate increase, and conduct decline, gated by available history. Each signal is converted to a within-faculty historical percentile using calibrators fit only on the training side of temporal validation. The domain score combines the strongest two positive deterioration percentiles.

## Structured routing
No free neural gate in Suite10. Expert eligibility is domain score times temporally validated expert reliability. Uncertainty does not choose an expert; it only caps specialist authority.

For expert k:
    a_ik proportional to m_ik * rho_fk
where rho_fk is a shrinkage estimate of historical in-domain validation reliability for faculty f.

Total specialist authority:
    A_i = domain_strength_i * reliability_strength_i * [a_min + (a_max-a_min) * uncertainty_i]
with uncertainty from Committee spread/entropy percentiles. NULL weight is 1-A_i. Specialist weights sum to A_i.

## Reliability
Expert reliability is estimated only from held-out historical temporal validation. Each faculty-domain reliability is shrunk toward a cross-faculty global reliability so small faculties do not overfit tiny validation subsets. Experts with insufficient validation support receive the global/shrunk estimate or zero if the global mechanism itself is unsupported.

## Training/selection
- Reproduce Suite7 using only the training side of the temporal split.
- Freeze shared encoder, faculty embeddings, adapters, and router.
- Train residual experts with semantic domain weights.
- Select each expert's epoch independently using future in-domain AP + F2@20 uplift versus Suite7/Committee anchor on temporal validation.
- Restore each expert to its own selected epoch.
- Estimate reliability from validation.
- For final refit, load the actual fitted Suite7 checkpoint, fit semantic calibrators on all historical OOF rows, train each expert for its selected epoch count, and route target rows with the frozen structured semantic policy.

## Primary outputs
1. Domain viability: rows, failures, prevalence, mean/core membership, Committee AP/AUC/F2.
2. Domain overlap/Jaccard matrices.
3. Train/validation/target domain drift.
4. Expert in-domain and out-of-domain AP/AUC/F2 uplift.
5. Expert specialization contrast (in-domain uplift minus out-of-domain uplift).
6. Faculty/global/shrunk reliability.
7. Routing authority/NULL/uncertainty diagnostics.
8. Residual correlations.
9. Committee -> Suite5 Shared -> Suite7 -> optional Suite8/Suite9 -> Suite10 comparisons.
10. Fixed and drift-aware operational metrics and full confusion matrices.
11. Per-row scores with all domain scores, specialist weights, residuals, authority, uncertainty and support regime.

## Success criteria
The experiment is scientifically successful only if several experts show positive held-out in-domain uplift, in-domain uplift exceeds out-of-domain uplift, domain membership is reasonably stable through time/faculties, structured routing captures those gains without label-conditioned responsibility, and the model does not rely on a high-entropy learned gate.
