# Suite16 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 Suite16 as a self-contained sequential conditional tree-boosting research package.

**Architecture:** Reuse Suite15 canonical/structured loading. Engineer three new chronology-safe specialist views, then train optional HGB/Newton and GRANDE-style correction stages sequentially from the immutable Committee under multi-window temporal validation.

**Tech Stack:** Python, NumPy, pandas, SciPy sparse SVD, scikit-learn HistGradientBoostingRegressor, PyTorch.

**Spec:** `docs/superpowers/specs/2026-08-30-suite16-design.md`

## Global Constraints
- Committee is the exact Stage-0 anchor.
- No gate, router, MoE authority, E0, or absolute target-semester feature.
- S2 contains dynamics only; S3 peer statistics use observations strictly earlier than each row semester.
- Every stage has exact no-op fallback.
- Holdout folds 6–7 never influence stage selection.

---

### Task 1: Specialist feature builders
- [ ] Write failing tests for S1/S2/S3 semantics and chronology.
- [ ] Implement latent factor cache and redesigned feature builders.
- [ ] Run focused tests.

### Task 2: Temporal windows and admission
- [ ] Write failing tests for contiguous prospective windows and robust admission.
- [ ] Implement window builder and stage-selection statistics.
- [ ] Run focused tests.

### Task 3: HGB Newton conditional booster
- [ ] Write failing tests for Newton residuals, exact no-op and positive correction learning.
- [ ] Implement HGB regressor stage and cross-fit/refit helpers.
- [ ] Run focused tests.

### Task 4: GRANDE-style differentiable booster
- [ ] Write failing tests for hard forward routing, straight-through gradients, bounded corrections and identity candidate.
- [ ] Implement self-contained hard axis-aligned ensemble.
- [ ] Run focused tests.

### Task 5: Sequential chains and diagnostics
- [ ] Write failing tests for skip/accept stage behavior and cross-fitted downstream bases.
- [ ] Implement HGB and GRANDE chains, single-stage audits, reporting, caches and CLI.
- [ ] Run end-to-end synthetic smoke.

### Task 6: Package verification
- [ ] Run full pytest, py_compile, shell syntax, CLI help and extracted-ZIP smoke.
- [ ] Verify no stale Suite15 user-facing names and package Suite16.
