Residual investigation canvas
Start a mission, inspect the residual field, then capture suspicious points or regions.
Mission setup
Score0
Streak0
Accuracy0%
Time--
Level1
Player1
Model metrics
| Metric | Current | After fix |
|---|
Before and after comparison
Model workshop
Experiment with model settings before retraining. Changes affect the simulated correction strength.
Case briefing
Each case hides a regression weakness. Random residual scatter is usually healthy. Structured shapes deserve investigation.
Learning summary
Complete a diagnosis to unlock a focused explanation and recommended next steps.
Player history
Achievements
First Find
Solve one case.
Solve one case.
Error Streak
Reach a three-case streak.
Reach a three-case streak.
Outlier Hunter
Catch an outlier case.
Catch an outlier case.
Variance Watch
Diagnose heteroscedasticity.
Diagnose heteroscedasticity.
Perfect Audit
Finish with no false alarms.
Finish with no false alarms.
Model Mechanic
Apply a successful repair.
Apply a successful repair.
Expert Detective
Solve an expert case.
Solve an expert case.
Battle Analyst
Complete a two-player round.
Complete a two-player round.
Accessibility and preferences
Diagnostic guide
Healthy residuals
Look for random scatter around zero, stable spread, and no obvious sequence or curve.
Bias and non-linearity
A shifted cloud suggests bias. A curved cloud suggests the model misses functional structure.
Changing variance
A funnel or fan shape indicates non-constant residual variance and may require transformation or weighting.
Outliers and leverage
Large residuals are response outliers. Extreme predictor positions create leverage and may become influential.