# Suite7 Shared Backbone + Faculty Adapter Design

## Research question
Does adding a small faculty-private representation adapter to the cross-faculty shared Residual-MoE reduce negative transfer on strong faculties while preserving gains on data-poor faculties?

## Controlled comparison
Suite7 keeps the Suite5 Committee, exact Suite5 per-fold common feature schema, faculty embedding, shared encoder, residual experts, NULL route, router, losses, sampling, chronological validation, and operational evaluation unchanged. The only architectural change is a faculty-specific residual adapter between the shared encoder and the experts/router.

## Architecture
For row x from faculty f:

1. Shared Suite5 path: `[x, faculty_embedding(f)] -> SharedEncoder -> h`.
2. Faculty adapter: `A_f(h) = W_up,f GELU(W_down,f h + b_down,f) + b_up,f`.
3. Private adaptation: `h_f = h + adapter_scale * A_f(h)`.
4. The same shared residual experts and router consume `h_f`.
5. The final Committee-anchored logit remains `logit(q) + sum_k w_k delta_k`.

Default adapter bottleneck is 16 for a 128-dimensional representation.

## Initialization
Do not zero-initialize the entire adapter. `W_down` and `b_down` use normal PyTorch initialization; only `W_up` and `b_up` are initialized to zero. Therefore `A_f(h)=0` for every faculty at initialization and Suite7 starts at the exact Suite5 function, while the output projection can learn immediately. Shared Suite5 modules are instantiated before adapters so the same seed reproduces Suite5 shared initialization exactly.

## Data and chronology
Reuse the completed Suite5 canonical cache. For each target fold, use the exact `common_raw_features` stored in Suite5 `fold_N/feature_schema.json`. Historical OOF rows from all faculties remain the training stream. The latest historical blocks per faculty remain validation. No holdout labels tune thresholds or model selection.

## Evaluation
Compare Committee, Suite5 shared Residual-MoE (read-only), and Suite7 shared+adapter on every faculty/fold. Report AP, AUC, exact warning-budget P/R/F1/F2 at 10/15/20/25%, development-frozen fixed threshold metrics, drift-aware metrics, and confusion matrices. Optional reference CSV adds original V2 and Suite4 Council.

## Adapter diagnostics
Report per faculty/fold/group: adapter L2 norm, representation L2 norm, adapter-to-shared relative norm, cosine(h, h_f), mean absolute adapter activation, NULL weight, routing entropy, max weight, correction magnitude, and residual correlations. Save per-faculty adapter parameter norms after each fold.

## Interpretation
Suite7 is successful if it improves macro/median AP and F2 over Suite5 shared while reducing negative transfer on IT/Business without erasing improvements on smaller faculties. Adapter magnitude alone is not evidence of useful adaptation; it must be linked to holdout metric gains.
