Score
0
Round
1
Streak
0
Accuracy
0%
Hints
3
Time
--
Challenge setup
Choose a task, model, explanation method, and difficulty.
Daily mode uses the date and seed for a repeatable challenge.
Prediction workspace
Current model prediction
0
Baseline
0
Positive
Negative
Sensitive or immutable
Your explanation
Drag features into rank order, choose direction, and estimate contribution.
Feedback and explanation
Compare your reasoning with the model’s local explanation.
Submit your explanation to receive ranking, direction, estimation, fairness, and counterfactual feedback.
Plain-language explanation
The explanation will appear here after submission or reveal.
What-if simulator
Change actionable values and watch the prediction update.
Counterfactual target
Find the smallest valid change that flips the outcome.
Distance cost0.00
Adjust actionable features to cross the decision threshold.
Explanation analytics
Inspect local contributions, global importance, confidence, and performance history.
| Round | Scenario | Score | Accuracy | Time |
|---|---|---|---|---|
| No completed rounds yet. | ||||
Data, classroom, and accessibility tools
Export progress, import a custom scenario, and adjust the interface.
Learning guide
Use explanations responsibly and distinguish evidence from causation.
A local contribution explains how one feature moves a specific prediction away from a baseline. Positive and negative signs describe direction, not moral value. Contributions can differ for another record.
Local importance concerns one prediction. Global importance summarises behaviour across many records. A feature can be globally important yet irrelevant for the current case.
A counterfactual finds a small change that would produce a different output. Useful counterfactuals should avoid immutable, sensitive, or unrealistic changes. They describe model behaviour rather than guaranteed real-world outcomes.
An explanation does not prove causation or fairness. Sensitive attributes and proxy variables require audits across groups. Confidence scores can also be misleading when data changes or the model is poorly calibrated.