# Cross-Faculty Residual MoE Design

## Goal
Build a Committee-anchored cross-faculty residual Mixture-of-Experts (MoE) that keeps each faculty's canonical RF/ET/HGB Committee fixed, learns shared residual specialists across faculties, jointly learns a NULL-aware router, and evaluates the system against the Committee plus optional external V2/Suite4 Council references using the same chronological fold protocol.

## Core prediction model
For faculty-specific Committee probability q(x), define z_q=logit(q). A shared bank of K residual specialists predicts bounded logit corrections delta_k(x). A router produces weights over a NULL expert and K residual experts. The NULL expert has delta_0=0. Final prediction is

z_final = z_q + sum_{k=1..K} w_k delta_k,
p_final = sigmoid(z_final).

The model is initialized so every residual output is zero, making the initial model exactly equal to the Committee.

## Inputs
The shared numeric representation is the exact raw-feature-name intersection across participating faculties plus V2-style Committee meta features and five chronology-safe support/history features. Faculty identity is supplied separately to a learned embedding. Scaling is fit on historical training rows only.

## Cross-faculty chronology
Each target fold loads canonical per-faculty caches. Historical specialist training uses only rows with chronology-safe cross-fitted Committee predictions from each faculty's fold training side. Target-fold labels from any faculty are never used in fitting. The final target prediction is produced for each faculty's fold test side.

## Joint specialization
For residual expert k, counterfactual probability is sigmoid(z_q + delta_k). Per-row utility is Committee binary log-loss minus counterfactual expert binary log-loss. NULL utility is zero. A detached softmax over [0,u_1,...,u_K] forms a soft capability target for the router. Router KL-to-capability is trained jointly with the final weighted prediction loss.

Prediction training uses faculty-balanced, class-aware sampling: each faculty receives equal expected sampling mass, while within each faculty effective-number class weights upweight failures without forcing identical prevalence.

A mild mutual-information-style router term encourages different residual experts to own different inputs conditional on using a non-NULL expert. A weighted residual-magnitude penalty bounds unnecessary Committee corrections.

## Temporal model selection
For each target fold, the latest N historical temporal blocks per faculty are held out as validation. The shared model is trained on earlier historical rows. Epoch selection uses a macro-faculty score: median faculty delta-AP plus a configurable weight times median faculty delta-F2@20 relative to the Committee. After selecting an epoch, the model and scaler are reinitialized and refit on all available historical OOF rows for exactly that many epochs before target prediction.

## Target-only ablation
An optional ablation trains the same architecture on selected target faculties alone while using the exact same common feature schema selected by the shared model. This isolates the value of cross-faculty transfer from architecture effects.

## Evaluation
For every faculty/fold, report Committee and residual-MoE AP/AUC plus F2 at 10%, 15%, 20%, and 25% warning budgets. For each faculty, tune each model's F2 threshold and warning-ratio policy only on development folds 1-5, then freeze those policies for holdout folds 6-7. Print V2-style fixed-threshold and drift-aware precision/recall/F1/F2/warning-rate/confusion matrices.

Across faculties, report macro mean/median/worst delta AP/AUC/F2@20 and win rates. Save per-fold scores, router weights, residual corrections, faculty/expert specialization summaries, thresholds, warning metrics, and macro summaries.

Optional standardized reference metrics may be loaded so the console/final CSV can include original V2 and Suite4 Council side-by-side without retraining those systems.

## Safety and scope
The canonical Committee is read-only. No target labels are used for training, validation, score alignment, threshold selection, or drift-rate estimation. V1 does not add a separate environment router, direct AP surrogate optimization, relational embeddings, retrieval memory, or hard routing; those remain later research directions.
