# Suite10 Semantic Specialists Implementation Plan

> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.

**Goal:** Build chronology-safe semantic academic-mechanism specialists on the Suite7 shared-private backbone.

**Architecture:** Four residual experts receive soft, observable domain responsibilities. Shared representation is frozen during specialist discovery. Structured semantic routing combines domain membership, temporal reliability, and Committee uncertainty as an authority cap.

**Tech Stack:** Python, NumPy, pandas, PyTorch, scikit-learn.

**Spec:** `SUITE10_DESIGN.md`

## Global Constraints
- Reuse existing Suite5 cache and Suite7 checkpoints; no Committee retraining.
- Labels must never enter domain membership or routing at target time.
- Validation calibrators must be fit only on the training side of chronological splits.
- Preserve exact Suite7 fallback as a separately reported safety output.
- Keep uncertainty as authority only, never as a semantic expert domain.

### Task 1: Semantic domain calibrator
Create `suite10_semantic_domains.py` and tests for student/course sparse, temporal deterioration, history-rich relational, per-faculty percentile fitting, and label independence.

### Task 2: Semantic residual training and routing
Create `suite10_semantic_moe.py` and tests for domain-weighted expert loss, per-expert gradient isolation, structured authority/NULL weights, and reliability shrinkage.

### Task 3: Diagnostics
Create `suite10_diagnostics.py` and tests for domain viability, overlap, drift, expert in/out-domain metrics, specialization contrast, residual correlation, and authority summaries.

### Task 4: End-to-end trainer
Create `train_suite10_semantic_specialists.py`, reusing Suite7 cache/checkpoints and reporting Committee/Shared/Suite7/Suite8/Suite9/Suite10 candidate/final metrics with temporal validation and operational evaluation.

### Task 5: Verification and documentation
Add synthetic smoke tests, README, compilation/CLI checks, and package `suite10_semantic_specialists.zip`.
