Lab
Evaluation Playground
Move the threshold and watch precision, recall and F1 change.
A classifier outputs a score, not a decision. You choose the threshold that turns scores into yes/no calls, and that single choice moves precision and recall in opposite directions. Drag the threshold and watch the confusion matrix and metrics respond.
Confusion matrix @ 0.50
Metrics across all thresholds
Raise the threshold and you demand more confidence before calling something positive: precision climbs (fewer false alarms) while recall falls (you miss borderline true cases). Lower it and the trade reverses. F1, the harmonic mean, peaks somewhere in the middle and is a common single-number summary when you care about both.
Honest note: this is one small, fixed dataset with clean labels. Real evaluation adds class imbalance, label noise, and the question of which errors actually cost you, a spam filter and a cancer screen sit at very different points on this curve.
Challenge
You are triaging urgent support tickets and missing an urgent one is far worse than a false alarm. Which metric do you optimise, and which way do you move the threshold? Find a setting that catches every positive, what does it cost in precision?