← Course
0/15

Module 01 · Foundations · Interactive lesson

Train a
Classifier

A classifier is just a line drawn through data. Here, 120 veterans wait to be sorted into a benefits fast-track or a manual-review queue. Move the line by hand, or let gradient descent train it — then watch what the most accurate line does to fairness.

Decision boundary

step 0

Scatter plot of 120 veterans by two features, with a movable decision boundary line.Horizontal axis is documentation completeness, vertical axis is record depth. Circles are Active-component veterans and squares are Reserve. Filled marks should be fast-tracked; outlined marks need manual review. A diagonal line is the model's current decision boundary; misclassified points are ringed in red.feature 1 · documentation →feature 2 · record depth →
Active (circle)Reserve (square)Fast-track (filled)Misclassified

Live metrics

Accuracy

72%

86/120 correct

Fairness gap

10%

within limit

Active · fast-tracked20%
Reserve · fast-tracked10%
Fairness guardrail: PASS

Tune the model

1.00
0.40
-0.90
01

What training is

Gradient descent nudges the three numbers — two weights and a bias — a little each step to reduce error. Press Auto-train and watch the boundary slide into place on its own.

02

Accuracy isn't fairness

Because Reserve records skew lower on documentation, the most accurate line fast-tracks Active veterans far more often. Accuracy climbs while the fairness gap widens past the limit.

03

Why governance

A model can be 'right' on average and still systematically disadvantage a group. The Guardrailed Classroom measures that gap and blocks the deploy until it's closed.

Why this matters

Optimizing for accuracy alone is not a neutral act. The same training run that makes the model more accurate can make it less fair— which is exactly why a scholar in the Guardrailed Classroom doesn't just train a model, they train it under a guardrail that watches the gap.

Mission first, people always.