# Suite17 — Robust Forward Complementary Specialist Boosting

## Research question
Can adaptive conditional boosting combine transferable relational/mastery evidence with faculty-local temporal evidence, preserving Suite16's cross-faculty gains while recovering Council's IT/Business advantage?

## Immutable anchor
The Committee is Stage 0 and is never retrained or replaced.

## Candidate specialist pool
- **R — Relational Compatibility**: chosen by chronology-safe ablation among relation-only, relation+curriculum, relation+latent, and full relation+curriculum+latent views. Only one R variant survives per target fold.
- **M — Relative Mastery**: preserve Suite16 peer-relative component/mastery phenotype.
- **L — Faculty-Local Temporal Tabular Evidence**: chronology-safe faculty-specific all-history and recent-history HGB predictors are treated as evidence. L consumes only their probabilities/logit/rank disagreements with Committee plus availability/support indicators; it does not consume raw flat features directly.
- **S2 Dynamics is removed** from the core specialist pool.

## Conditional learner
HGB/Newton conditional correction remains the primary learner. Homemade GRANDE is removed. Optional external differentiable-tree experiments are out of the main Suite17 decision path.

## Forward selection
Starting from Committee, all remaining specialists are evaluated conditionally on the current dense OOF score ledger. Select the robustly best candidate. Repeat for at most two deployable stages by default; a third stage is diagnostic-only unless it passes the stronger third-stage admission threshold.

## Dense chronology-safe ledger
Earlier-stage cross-fitted predictions never shrink the row universe. Rows lacking sufficient prior history use exact identity fallback (delta=0), with a coverage/availability indicator in later-stage controls. Downstream stages therefore receive 100% score-ledger coverage.

## Temporal admission
Each temporal window reports faculty median and mean delta AP separately. Admission requires positive global mean/median, adequate faculty win rate, non-negative recent-window mean and positive recent-window median, and no materially negative mean in the two newest windows. F1/F2 downside guards remain secondary.

## Correction authority
Use eta grid {0,.05,.10,.20,.30} and tighter bounded corrections. The objective is incremental evidence, not replacement of Committee.

## Rank-preserving alignment
After the final selected chain, align final logit location/scale to the target Committee logit distribution using unlabeled target scores only. Positive affine logit alignment preserves AP/AUC ordering while reducing score-scale drift for historical thresholds.

## Required diagnostics
- R ablation historical selection and target counterfactuals.
- L local-global/local-recent coverage, AP vs Committee, disagreement magnitude, and faculty-specific gains.
- Standalone R/M/L conditional audits.
- Every forward-selection candidate at every stage, selected candidate, rejection reason, window mean/median/worst statistics.
- Dense-ledger coverage and identity-fallback rates.
- Conditional correction magnitude, saturation, residual correlation, and target faculty effects.
- Raw vs aligned final ranking/calibration metrics.
- Complete comparisons with Committee, Suite16, Suite14, Council, V2 and earlier stored lineages.
- Council overlap table across available faculties, equal-faculty and row-size-weighted summaries.
