Make every prediction
explainable.
Inspect model reasoning, verify feature influence, find actionable counterfactuals, and audit subgroup behavior in one transparent, reproducible workspace.
Command Center
One observable prediction with a complete audit trail.
What drives this prediction?
Positive drivers increase risk, negative drivers reduce it. Contributions add exactly to the model log-odds.
Adjust model signals
Local Explanations
Trace every feature to the predicted result.
Signed contribution breakdown
Feature response
Complete model formula
Every feature contributes its coefficient times a centered, scaled input. The logistic function converts total log-odds into probability.
Global Importance
Contrast out-of-sample permutation impact with feature dependence.
Change in Brier loss
Partial dependence + ICE
Counterfactual Lab
Discover the smallest modeled changes that meet a target risk.
Risk reduction planner
Searches changes to one or two actionable inputs. Immutable signals are never modified.
Recommended scenario to test
Group Diagnostics
Inspect subgroup performance differences without hiding sample sizes.
Evaluation by group
| Group | Records | Actual rate | Mean risk | Brier ↓ | ROC AUC ↑ | Recall |
|---|
Decision threshold 50%
Changes confusion-matrix statistics and recall. It does not retrain the model or change its predicted probabilities.
Data & Methods
Bring a labeled dataset, inspect the methodology, and export an auditable report.
Import evaluation CSV
Load your own synthetic or appropriately authorized records. Files stay in this browser. The fixed illustrative model is not trained on imported data.
Export research artifacts
Includes model formula, local attributions, counterfactual plan, global permutation importance, group performance and data-source caveats.
Transparent by design
Additive logistic model with fixed published coefficients and direct log-odds decomposition.
ROC AUC and Brier score on labeled records; undefined metrics are displayed as N/A.
Seeded repeated feature permutation and mean loss difference relative to the unchanged model.
Discrete grid search of 1-2 editable features. Search is approximate, not a guaranteed continuous global optimum.
No server-side upload: all computations occur inside the browser session. Avoid entering sensitive data on public devices.
Demo models and labels are synthetic; scores are not suitable for operational, medical, hiring, or credit decisions.