# Suite9 — Competitive Responsibility Residual MoE

## Goal
Test whether true residual specialists emerge when experts receive different **training responsibilities**, rather than merely different feature views.

## Controlled baseline
Suite7 is the deployable architectural baseline. Suite8 is a diagnostic reference. Suite9 loads the fitted Suite7 checkpoint for each fold, preserves its faculty embedding, shared encoder, faculty adapters, and residual experts as the initialization, then adds Suite8-style latent expert views and changes only how specialists and the router are trained.

## Core hypothesis
Suite8 showed that latent ownership can create different views, but stronger diversity reduced forward AP/F2. This suggests that input diversity is not sufficient. Suite9 therefore creates sparse row-level responsibilities from each expert's counterfactual improvement over the frozen Committee and uses those responsibilities to train experts competitively.

## Phase 1 — competitive specialist discovery
For row i and expert k:

- Committee loss: l_i0 = BCE(y_i, q_i)
- Expert counterfactual: p_ik = sigmoid(logit(q_i) + delta_ik)
- Expert loss: l_ik = BCE(y_i, p_ik)
- Utility: u_ik = l_i0 - l_ik

NULL has utility 0. Sparse responsibilities are produced by a top-k softmax over [NULL, experts], with experts eligible only when their raw utility exceeds `min_expert_utility`. Responsibilities are detached before expert updates.

Only the residual experts and latent-view parameters train during Phase 1. The Suite7 faculty embedding, shared encoder, faculty adapters, and router remain frozen. Expert loss is the responsibility-weighted counterfactual BCE plus residual-magnitude and mild view-balance/specialization regularizers.

## Phase 2 — router distillation
Freeze representation, adapters, latent views, and experts. Train only the router to predict the frozen competitive responsibility targets. Router training uses KL divergence between learned routing weights and sparse responsibility targets.

Router model selection is deployment-oriented and compares the learned-router candidate directly against the exact Suite7 score on temporal validation. If no candidate exceeds Suite7 by `min_selection_improvement`, Suite9 falls back exactly to Suite7 for that fold.

## Phase 3 — optional joint refinement
Disabled by default. If enabled, specialists and router are refined jointly at a much smaller learning rate using detached competitive responsibilities. The refined model is accepted only if its temporal validation score exceeds both the Phase-2 model and Suite7.

## Final prediction
z_final = logit(q) + sum_k w_ik * delta_ik

NULL is represented by router weight w_i0 and contributes zero correction.

## Required diagnostics
Per fold/faculty/group, report:
- NULL and expert responsibility mass
- responsibility entropy and top-1/top-2 assignment fractions
- responsibility mass for failures/normals/history regimes
- per-expert utility on assigned vs unassigned rows
- per-expert AP/AUC within assigned regions when defined
- residual correlation and expert-view cosine
- router entropy, max weight, NULL weight
- router-vs-responsibility top-1 accuracy and KL
- oracle-responsibility AP/F2 vs learned-router AP/F2
- Suite7 vs Suite9 candidate vs Suite9 deployed fallback
- rescued/lost failures and added/removed false positives at 20% warning budget
- frozen fixed and drift-aware operational metrics

## Baseline references
Each target score file must include Committee, Suite5 Shared, Suite7 Adapter, Suite8 True Specialist (if supplied), Suite9 learned-router candidate, oracle-responsibility diagnostic score, and Suite9 deployed score.

## Safety and chronology
All model selection uses historical/validation labels only. Holdout labels are never used to select specialist epochs, router epochs, fallback, thresholds, or warning ratios. Oracle-responsibility metrics are diagnostic only and never drive holdout deployment decisions.
