0.02 stochastic variation
Live laboratory
Probe the hidden decision process
Awaiting probe
Score
0
Rookie Investigator
Experiments
0 / 28
Efficient testing earns more
Prediction shift
0.000
From the baseline case
Confidence
—
May be hidden by difficulty
Input laboratory
Change one clue or combine several
Diagnostics
Observed model behaviour
Evidence log
Experiment history
| # | Action | Changed feature | Prediction | Shift | Confidence |
|---|---|---|---|---|---|
| Run an experiment to create evidence. | |||||
Case objectives
Investigation missions
Case completion0%
Investigator tools
Focused tests
Notebook
Hypothesis and findings
Guidance
Hints and explanations
Change one feature first. Then compare prediction direction, size, and confidence.
Live console
Investigation events
Final report
Investigation summary
Complete missions and submit a conclusion to generate your investigation report.
Learning guide
How black-box investigation works
Perturbation testing
Change inputs deliberately while holding others stable. The resulting output movement reveals local model behaviour without exposing internal parameters.
Counterfactual reasoning
Search for the smallest realistic change that flips a class or reaches a target. Counterfactuals explain actionable decision boundaries.
Robustness and fairness
Test small noise, missing data, extreme values, and protected attributes. Stable, equitable behaviour strengthens confidence in deployed predictions.