Classifier Trade-offs Lab

Move the threshold and watch trade-offs appear

This experiment uses a fixed set of twelve fictional observations. Each has an invented score between zero and one hundred and a known positive or negative label. A score at or above the chosen threshold counts as a positive prediction.

Change the slider and watch the true-positive, false-positive, true-negative and false-negative counts update. Precision describes how many predicted positives are actually positive; recall describes how many actual positives were detected. Accuracy is the share of all correct predictions.

Make a decision, then explain it

At a low threshold the model may flag almost everything, producing high recall but many false alarms. At a higher threshold, fewer cases are flagged, but some true cases may be missed. Neither threshold is automatically the correct choice without information about the consequences of those errors.

Try the extremes. If the classifier never predicts positive, precision has a zero denominator and is undefined rather than magically perfect. This distinction matters when dashboards present rounded statistics without revealing the underlying counts.

What the exercise cannot establish

The data are made up, so the experiment does not measure an actual AI product, predict future outcomes, or test generalization. Real performance needs independent representative testing and attention to bias, changing conditions and harms.

Read the Google classification-metrics guide and NIST AI risk material below for definitions and broader considerations.