HUGIML Governance StudioStatic interactive preview powered by demo artifacts
Interface
HUGIML Governance StudioOcean
Workbench & Governance Dashboard
Configure and compare models in Workbench, then promote a selected HUGIML run for governance review.
Data profilerPer-model configurationComplexity evidence
HUGIML Governance Studio
Static interactive preview powered by demo artifacts
Experiment Workbench
Model Comparison & Benchmarking
Configure HUGIML with a named grid and downstream estimator, compare validation evidence, and inspect the exact final representation before promoting a fitted run into Governance.
1Select modelsChoose HUGIML and comparison models.
2Configure candidatesSelect performance, performance_ho, or interpretability.
3Inspect representationSeparate RPTE leaves from direct source terms.
4PromoteCarry the same fitted model and validation evidence into Governance.
Experiments complete. Results refreshed for the selected demo dataset.
1. Select models
Baselines
Logistic Regression and Decision Tree.
Ensembles
Random Forest, XGBoost, and LightGBM when available.
Interpretable models
HUGIML is governance-native; EBM and RuleFit provide comparison artifacts.
2. Configure HUGIML
Named search grid
parameter
guided values
Execution contract
RPTE final LR columns: all RPTE leaf indicators first, followed only by supplied HUGIML source columns that were not used in an accepted RPTE tree split.
Leaves root-to-leaf conjunctionsDirect original features direct LR termsDirect HUG patterns direct LR termsDirect pairs direct LR terms
Source columns used by a tree split are represented through leaf paths and are not duplicated as standalone LR terms.
Leaderboard ranks candidate runs by mean-fold CV ROC-AUC. Pooled out-of-fold diagnostics are kept as a separate validation summary.
run_id
model
grid
downstream
best
mean CV ROC-AUC
pooled OOF ROC-AUC
F1
fit sec
Curves use stitched out-of-fold probabilities from the same evaluation bundle carried into Governance.
ROC comparison
HUGIML RPTEHUGIML LRRFLR baseline
Precision-recall comparison
HUGIML RPTEHUGIML LRRFLR baseline
RPTE interpretation shows final LR terms only: leaf indicators and direct source terms. Source columns used in accepted splits appear inside the flat-tree paths.
Choose a fitted-model artifact. Direct source tables are filtered to direct terms only.
Model drill-down links the selected validation run to the fitted representation that will be promoted.
HUGIML Governance Studio
Audit & Evidence Dashboard
Governance separates RPTE construction inputs, accepted tree splits, leaf-rule indicators, and direct LR terms so that representation evidence and governance actions match the fitted model.
1ValidateReview the promoted out-of-fold evidence.
2Trace representationFollow source inputs through tree use to final LR terms.
3Review actionsSeparate direct-term review from a full pipeline rebuild.
No promoted HUGIML run yet. Go to Workbench → Results → Model drill-down → Promote.
Overview
Executive model card for the promoted fitted run and its final representation.
Column roles
Selected HUGIML parameters
Representation status
Evidence inventory
artifact
status
scope
Validation
Governance displays the same promoted out-of-fold evidence used in Workbench. Mean-fold and pooled OOF ROC-AUC are labeled separately.
Confusion matrix — pooled OOF
Fold stability
CV candidate evidence
grid
downstream
fold
ROC-AUC
status
Representation Audit
The RPTE path is audited as a four-stage flow. Final LR evidence contains leaf rules plus only direct original features, HUG patterns, and augmented pairs.
Tree-used source families
Final LR term families
A source column used in any accepted RPTE split is represented through root-to-leaf conjunctions. It is not repeated as a standalone LR term. Only the remaining source columns enter the direct-term block.
Leaf coefficients
leaf
coefficient
odds multiplier
support
final LR term
family
coefficient
odds multiplier
support / availability
role
Adaptive binning
Training-only bin selection supplies original and pattern candidates to RPTE.
source
chosen B
IG
RPTE role
Interaction-relaxed sources
Survivor evidence changes mining eligibility; it does not by itself define a final LR term.
source
best partner
interaction score
downstream role
Pattern Inventory
Patterns are separated by fitted role: tree-used patterns appear inside RPTE paths; direct HUG patterns remain direct LR terms with their own coefficients.
These patterns contribute through leaf conjunctions. No standalone pattern coefficient is shown.
pattern
trees
split count
support
final role
These patterns were supplied to RPTE but were not used in an accepted split; they are direct LR terms.
pattern
coefficient
odds multiplier
support
final role
Leaf activation coverage
Interpretation boundary
Rows with no active leaf in one tree still receive exactly one leaf indicator from every other accepted tree, plus values from any direct source terms. Coverage should therefore be reviewed by tree and by direct-term family rather than through HUG-pattern activation alone.
Case Review
Case contributions are grouped into active RPTE leaves and active direct source terms.
Data Quality & Policy
Sensitive and proxy review distinguishes tree-path use from direct-term use.
Missing-value handling
source
missing %
RPTE role
policy action
Sensitive / proxy representation
field
tree-path use
direct source terms
review
Configuration Comparison
Configurations include the downstream estimator and final representation structure.
grid
downstream
mean CV ROC-AUC
leaf terms
direct original features
direct HUG patterns
direct pairs
note
Representation Pruning
RPTE governance uses separate action scopes. Direct-term analysis applies only to direct source terms; raw-input exclusion requires a full mining → RPTE → leaf + direct-source → LR rebuild.
Direct-term review
Removing selected direct terms means refitting the downstream LR on the unchanged leaf matrix plus the retained direct-term block.
term
family
coefficient
review status
Raw-input exclusion
Removing an original input can change mining, augmented-pair generation, accepted RPTE splits, the leaf inventory, and which source columns remain direct terms. It is therefore a complete pipeline rebuild, not direct LR pruning.
Leaf-rule review boundary
Leaf indicators are the RPTE tree output. A leaf can be reviewed by coefficient, support, path length, or sensitive adjacency. Removing a leaf term is a separate downstream refit decision and must not be presented as removing the original features or patterns found inside its path.
Monitoring
Monitoring is split into RPTE leaf activation stability, direct-term drift, and source-to-tree representation changes after retraining.