# Suite11 — Opportunity-Aware Latent Domains

## Research hypothesis
Suite9 showed that competitive responsibility can create genuinely different expert functions, while Suite10 showed that clear semantic domains do not guarantee useful specialization. Suite11 tests whether recurring **historical OOF correction opportunities** can be compressed into latent domains that are both predictively useful and prospectively identifiable from pre-outcome information.

## Non-negotiable chronology
- Committee/base-expert probabilities used for discovery are historical OOF probabilities from the existing cross-faculty cache.
- Outcome labels may be used only to define historical opportunity signatures and specialist training responsibilities.
- The target fold outcome is never used for discovery, routing, model selection, threshold tuning, or deployment decisions.
- Target hindsight domains are computed only after scoring for diagnostics.

## Pipeline
1. Reproduce the Suite7 training-side baseline using the same historical split.
2. Build historical OOF opportunity signatures from the three committee experts and committee score. Features include per-expert utility, committee BCE, signed error, missed-failure severity, false-alarm severity, committee rank, disagreement, utility dispersion, and entropy.
3. Define a NULL region as the lower-opportunity portion within each faculty/temporal block. Cluster only the higher-opportunity rows using MiniBatchKMeans. Distances are converted to soft latent responsibilities.
4. Train a prospective MLP router using only observable Suite5/Suite7 input features to predict NULL + latent domain labels. Validate router balanced accuracy, macro-F1, NMI, confusion matrix, and temporal prevalence shift.
5. Initialize residual specialists from the Suite7 checkpoint/backbone, freeze the shared representation, and train each residual expert using its soft latent-domain responsibility.
6. Fuse the committee with router-weighted specialist residuals. A chronology-selected authority scalar alpha shrinks all non-NULL router mass back toward the committee.
7. Deploy Suite11 only if its newest historical selection block improves the exact Suite7 reference according to median cross-faculty AP + F2@20; otherwise use the exact Suite7 score for that fold.

## Success criteria
Primary: substantial, broad improvement in forward holdout AP, Precision/Recall/F1/F2 and confusion-matrix TP/FN/FP/TN relative to V2, Suite4 Council, Suite7, Suite9 and Suite10 when those references are available.

Diagnostics are secondary: router predictability, domain stability, expert in-domain uplift, oracle gap and residual diversity. Low residual correlation alone is not success.
