# Suite18 — Multiscale FMM Representation + Robust Forward Specialist Boosting

## Research question

Does a fuzzy hyperbox representation add chronology-stable geometric information
to Suite17's semantically specialized R/M/L views without changing the Committee,
the specialist learner, or the forward-selection policy?

## Controlled change

Suite18 reuses Suite17 exactly for:

- immutable Committee Stage 0;
- R ablation family;
- Relative Mastery M;
- faculty-local L evidence construction;
- Newton-HGB conditional correction;
- temporal meta-window selection;
- admission guards;
- dense identity-filled OOF ledger;
- eta grid and correction bound;
- at most two deployable stages;
- rank-preserving final alignment;
- operational evaluation.

Only the representation supplied to each R/M/L HGB specialist changes:

    Suite17: raw specialist view
    Suite18: [raw specialist view | 46 FMM descriptors]

Each R ablation owns an independent FMM encoder.

## FMM preprocessing

At the first temporal block with sufficient strict prior history:

1. median imputation is fit on prior rows only;
2. ANOVA F-score selects top 32 raw view features;
3. min/max scaling is fit on that same prefix;
4. the preprocessing map is frozen for the rest of the fold.

Freezing is intentional.  Re-fitting min/max at every block would change the
coordinate system of existing hyperboxes; fitting on all historical rows would
leak future historical blocks into earlier descriptors.

All transformed future values are clipped to [0,1].

Defaults:

- top_k = 32
- theta = {0.15, 0.30, 0.50}
- gamma = 1.0
- confidence k = 0.5
- minimum strict-prior rows = 5000
- minimum examples in each class = 50

## Prequential chronology

For historical block b:

    query descriptor with FMM state trained on blocks < b
    update confidence statistics using labels from b
    absorb rows from b into the hyperboxes

Therefore current-block labels cannot influence their own descriptors.

Rows before the FMM initialization point receive 46 zeros.  Raw Suite17 features
remain present, so this is an exact identity fallback rather than row removal.

## 46 FMM descriptors

Per theta (14 features):

1. normal membership
2. failure membership
3. normalized failure membership
4. failure-minus-normal membership
5. maximum membership
6. top-class membership gap
7. winning class is failure
8. winning hyperbox confidence factor
9. log1p winning support
10. centroid L1 distance
11. centroid L2 distance
12. mean hyperbox width
13. log-volume summary
14. inside-winning-hyperbox indicator

Across three theta values:

15. failure evidence mean
16. failure evidence std
17. failure evidence max
18. scale agreement

Total = 3*14 + 4 = 46.

## Hyperbox confidence factor

The paper definition is retained:

    CF_j = (1-k) U_j + k A_j

U_j = usage count divided by maximum usage among hyperboxes of the same class.
A_j = correct-prediction count divided by maximum correct count among hyperboxes
of the same class.

Usage/correctness are updated only after a historical block has been queried.

## Common/Committee FMM

A separate FMM encoder is trained from the common flat Suite17 view.

It does NOT alter Committee q and is never a forward-selection candidate.
It exists for representation diagnostics and the paper-inspired hybrid branch.

## Paper-inspired diagnostic branch

Using the medium theta (default 0.30):

    final historical FMM hyperboxes
      -> hyperbox centroids as tree training samples
      -> confidence factor as sample_weight
      -> CART and Random Forest

Reported models:

- fmm_common_standalone
- fmm_common_cart_diagnostic
- fmm_common_rf_diagnostic

These scores never enter R/M/L selection or the deployed chain.

## Primary comparisons

When --suite17_results_dir is supplied, Suite18 adds the original
suite17_hgb_final score as `suite17_hgb_final_reference`, allowing the exact same
Suite17 evaluator to report direct holdout comparisons.

Important outputs:

- suite18_run_config.json
- fold_metrics.csv
- mean_fold_holdout.csv
- operational_metrics.csv
- primary_holdout_comparison.csv
- suite18_vs_suite17.csv
- suite18_vs_suite17_specialist_audit.csv
- fmm_feature_selection.csv
- fmm_prequential_audit.csv
- fmm_hyperbox_summary.csv
- fmm_view_summary.csv
- fmm_paper_hybrid_diagnostics.csv
