# Suite16 Sequential Conditional Gradient-Tree Boosting — Design

## Goal
Test whether redesigned, anti-redundant structured specialists can deliver substantial cumulative gains beyond the immutable RF/ET/HGB Committee.

## Immutable anchor
Stage 0 is the Committee mean probability. No neural E0 or global correction is used.

## Specialists
1. S1 Student–Curriculum Compatibility: prior student-course relation evidence, student-conditioned prerequisite readiness, course-relative ability gaps, and chronology-safe latent student×course factors.
2. S2 Student Dynamics: deltas, slopes, acceleration, volatility, drawdown and recent-vs-long-run changes only; absolute academic levels are excluded.
3. S3 Relative Mastery Phenotype: prior component scores normalized against prior-only course×component peer distributions, with type-specific, recency-weighted and target-related summaries.

## Conditional boosting
Every stage learns only a correction to the frozen previous logit. HGB uses a Newton pseudo-residual regressor. The differentiable model is a self-contained GRANDE-style hard axis-aligned tree ensemble with straight-through feature/split decisions and instance-wise leaf weighting. Every stage has an exact no-op candidate and is admitted only by chronology-safe temporal meta-validation.

## Temporal protocol
Create contiguous walk-forward windows that satisfy minimum faculty/failure/row support. Use the latest three qualifying windows for robust stage selection; generate cross-fitted stage predictions over all qualifying windows for downstream stages. Raw absolute semester is never a model feature.

## Evaluation
Primary AP; secondary AUC, F1, F2 and F2@20. Print single-stage redesigned-view audits, per-window stage deltas, admission reasons, cross-fit coverage, cumulative stage metrics, HGB-vs-GRANDE comparison, correction diagnostics, feature usage, faculty/fold tables, and comparisons to Committee plus saved earlier reference score files where available.
