# Suite17 — Robust Forward Complementary Specialist Boosting

Suite17 is the post-Suite16 redesign. It keeps Committee immutable and removes both the weak S2 Dynamics specialist and the homemade GRANDE branch.

## Candidate specialists
- **R**: one history-selected relational variant from `R_RELATION`, `R_RELATION_CURRICULUM`, `R_RELATION_LATENT`, `R_FULL`.
- **M**: Suite16 Relative Mastery specialist retained.
- **L**: new faculty-local all-history/recent HGB alternative-prediction evidence. L receives only predictions/disagreement/rank/support evidence, never raw flat features directly.

## Forward selection
At each stage, all remaining R/M/L candidates are evaluated conditionally on the current Committee-derived score using strict multi-window temporal validation. The best admitted candidate is deployed. The default deployable chain is capped at two specialists; the third is evaluated diagnostically but not deployed unless `--deploy_third_stage` is explicitly supplied.

Historical score ledgers are always dense. Rows without sufficient prior data receive exact identity fallback and an availability indicator; the row universe never shrinks stage by stage.

## Calibration
`suite17_hgb_raw` is the raw conditional-boosting score. `suite17_hgb_final` applies positive-affine per-faculty target-logit alignment to the Committee score distribution. The alignment uses no target labels and preserves AP/AUC ranking exactly.

## Run
```bash
cd /root/disk_3tb/earlywarning/30072026/suite17
bash run_suite17.sh
```

Default log:
`/root/disk_3tb/earlywarning/30072026/log/suite17.log`

Default results:
`/root/disk_3tb/earlywarning/30072026/results/suite17_robust_forward_specialist_boosting`

## Key result files
- `primary_holdout_comparison.csv`
- `faculty_head_to_head.csv`
- `suite17_vs_council_faculty.csv`
- `suite17_vs_council_summary.csv`
- `standalone_specialist_audit.csv`
- `standalone_specialist_audit_faculty.csv`
- `forward_candidate_audit.csv`
- `forward_stage_decisions.csv`
- `fold_selection_summary.csv`
- `correction_diagnostics.csv`
- `feature_usage.csv`
- `alignment.csv`

Each `fold_N/` also contains per-faculty score files, R-ablation/forward-selection diagnostics, local-evidence diagnostics, and selected specialist order.
