SUITE7 — CROSS-FACULTY SHARED RESIDUAL MoE + FACULTY-PRIVATE ADAPTER ==================================================================== Research question ----------------- Does a small private faculty adapter reduce the negative transfer observed in the fully shared Suite5 Residual-MoE while retaining useful cross-faculty transfer? Suite7 is a controlled Experiment B. The Committee, Suite5 common feature schema, shared encoder architecture, residual experts, NULL expert, router, loss, temporal validation, sampling, and operational evaluation remain unchanged. The only new representation component is: h = SharedEncoder([x, faculty_embedding]) h_f = h + A_f(h) experts/router consume h_f where A_f is a faculty-specific bottleneck adapter. Default: 128 -> 16 -> 128 Initialization -------------- Do NOT initialize the complete adapter to zero. The down projection keeps normal PyTorch initialization. Only the up projection weight and bias are initialized to zero. Therefore A_f(h)=0 exactly at initialization and h_f=h, so Suite7 begins at the same function as Suite5. Because Suite5's residual experts also start with zero output layers, the first minibatch primarily activates the shared residual path; the adapter up projection then begins learning and the down projection follows once the up projection is non-zero. The regression tests verify that this path wakes up rather than remaining dead. The Suite5-equivalent faculty embedding, encoder, residual experts, and router are constructed BEFORE the adapters. With the same random seed, their initial random parameters therefore match Suite5's initialization sequence. Runtime / cache --------------- NO new Committee cache is needed. Reuse: /root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe/cache Suite7 also reads the completed Suite5 shared results as a read-only reference: /root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe It does not retrain Suite5. Only one Suite7 shared+adapter model is trained per fold. Recommended full run -------------------- cd /root/disk_3tb/earlywarning/30072026/suite7 mkdir -p /root/disk_3tb/earlywarning/30072026/results/suite7_shared_private_adapter mkdir -p /root/disk_3tb/earlywarning/30072026/log python3 train_suite7_shared_private_adapter.py \ --cache_dir /root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe/cache \ --shared_results_dir /root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe \ --models_dir /root/disk_3tb/earlywarning/30072026/results/suite7_shared_private_adapter \ --faculties all \ --dev_folds 1,2,3,4,5 \ --holdout_folds 6,7 \ --seed 42 \ --device auto \ --n_experts 4 \ --faculty_dim 8 \ --encoder_hidden 256 \ --representation_dim 128 \ --expert_hidden 64 \ --delta_max 2.0 \ --dropout 0.10 \ --router_temperature 1.0 \ --adapter_dim 16 \ --adapter_dropout 0.0 \ --adapter_scale 1.0 \ --epochs 60 \ --min_epochs 8 \ --patience 8 \ --fallback_epochs 20 \ --val_blocks 2 \ --min_train_rows 800 \ --min_val_rows 200 \ --batch_size 1024 \ --predict_batch_size 4096 \ --lr 0.001 \ --weight_decay 0.0001 \ --grad_clip 5.0 \ --print_every 5 \ --router_lambda 0.5 \ --specialization_lambda 0.01 \ --magnitude_lambda 0.01 \ --utility_temperature 0.25 \ --sampling_beta 0.9999 \ --macro_f2_weight 1.0 \ --min_selection_improvement 0.00001 \ --threshold_beta 2.0 \ --min_precision 0.15 \ --max_warning_ratio 2.5 \ --min_flags 5 \ > /root/disk_3tb/earlywarning/30072026/log/suite7_shared_private_adapter.log 2>&1 Monitor: tail -f /root/disk_3tb/earlywarning/30072026/log/suite7_shared_private_adapter.log Main outputs ------------ suite7_fold_metrics.csv Per-fold Committee / Suite5 Shared / Suite7 Shared+Adapter AP, AUC, and P/R/F1/F2 at 10/15/20/25% warning budgets. suite7_holdout_pooled.csv Faculty-wise pooled holdout ranking/budget metrics. suite7_warning_metrics.csv suite7_threshold_policies.csv Separate development-frozen fixed and drift-aware operational policies for all three models, including precision/recall/F1/F2/rate and confusion matrices. suite7_shared_private_comparison.csv Main causal comparison. Includes Adapter-vs-Shared deltas for AP, AUC, F2@20, fixed F2 and drift F2, plus descriptive ADAPTER_STRONGER / SHARED_STRONGER / MIXED evidence labels. suite7_adapter_diagnostics.csv Per faculty/fold and all/failure/normal/history groups: adapter norm, relative adapter norm, cosine(h,h_f), mean absolute adapter activation, NULL weight, router entropy, max routing weight, correction magnitude and expert statistics. suite7_holdout_adapter_router_summary.csv Compact holdout summary comparing Suite7 router behavior to Suite5 shared behavior and exposing adapter strength by faculty. suite7_adapter_parameter_norms.csv Down/up projection parameter norms for every faculty adapter and fold. suite7_residual_correlations.csv Pairwise residual-expert correlations after private adaptation. fold_N/shared_private_model.pt fold_N/shared_private_training_history.csv fold_N/_scores.csv Row-level scores, router weights, residuals and adapter diagnostics. Optional V2 / Council references -------------------------------- Use collect_reference_metrics.py in this folder to create a reference CSV, then pass: --reference_metrics /path/to/reference.csv Suite7 will additionally write: suite7_official_comparison.csv with historical V2 / competence_council metrics merged into the main faculty table. Interpretation target --------------------- The key hypothesis is NOT merely that adapters improve mean AP. We want to see: 1. IT / Business recover from Suite5 negative transfer; 2. smaller faculties such as German/Japanese/Chinese/Spanish retain shared gains; 3. macro/median Adapter-vs-Shared AP and F2 become positive; 4. adapter norms differ meaningfully across faculties rather than every faculty receiving the same-sized perturbation; 5. operational recall/F2 do not inherit Suite5's systematic conservative shift. If adapters help only by making large uncontrolled representation shifts, or if the same adapter magnitude appears everywhere, that is weak evidence for the shared- private hypothesis.