Laboratory / Computing
Learning systems and evaluation
Move a classification threshold and watch false positives and false negatives trade places. Then change class prevalence and observe why precision changes even when sensitivity and specificity stay fixed.
What changed
Threshold 55% and prevalence 35% reshape the confusion matrix. Precision can change even when the detector itself does not.
Return to the route
Experiment after prediction.
Before changing a control, say which visible feature should move and which should remain invariant. That prediction is the useful part of the experiment; motion alone is not evidence.