One running question across every arc: the quality team wants a flag that lights up on admission when a patient is likely to be back within 30 days — which model would you actually trust to raise that flag, and how would you know? Opens with a one-line model that scores 88.84% accuracy and asks whether that's an argument for deploying it. Arc 1 builds the module's patient-safe split (reopening Module 02's independence lesson and its own fishing trip, with a model this time). Arc 2 extends Module 02's r into a fitted line, then bends that line through a sigmoid into logistic regression. Arc 3 builds the confusion matrix and shows why 0.5 is a choice, not a law. Arc 4 grows a decision tree by hand (Gini, then depth) and reopens Arc 1's sealed box. Arc 5 shows which model families actually care about feature scale (a live three-way experiment), then finds groups with k-means, label held out, checked honestly after the fact.
Builds forward from Modules 01-02 — tidy data, descriptive statistics, correlation, and hypothesis testing are assumed, not re-taught. The dataset, split seeds, and several exact numbers (r=0.4661, the 40-visit patient, the 300-visit fishing-trip sample) are reused verbatim from Module 02.
This module's model-fitting cells and the project starter use scikit-learn, which wasn't in the
toolchain before this module. Re-run env/setup.sh once before starting — it installs the
newly pinned dependency and re-runs the smoke test.