Mastery: ship checklist

Shipping means handing off evidence, not only source code. Publish the depth, leaf constraints, feature order, class meanings, held-out metrics, and a text or visual representation. State that paths explain model mechanics, not causation. Rebuild or rerun from the documented starting point. If another person needs an undocumented fact from your machine, the handoff is incomplete.

Keep the example small enough to inspect manually. Small does not mean careless: boundary values, file locations, feature order, and held-out data still determine whether the result means what you claim.

from sklearn.tree import DecisionTreeClassifier, export_text

X = [[1, 0], [2, 1], [4, 1], [5, 3], [6, 2], [7, 4]]
y = ["needs-practice", "needs-practice", "almost", "ready", "ready", "ready"]

tree = DecisionTreeClassifier(max_depth=2, min_samples_leaf=1, random_state=7)
tree.fit(X[:5], y[:5])
print(tree.predict([[7, 4]])[0])
print(export_text(tree, feature_names=["hours", "practice_tests"]))

Expected evidence:

ready
|--- hours <= 3.00
|   |--- class: needs-practice
|--- hours >  3.00
|   |--- hours <= 4.50
|   |   |--- class: almost
|   |--- hours >  4.50
|   |   |--- class: ready

The output may include version-specific details such as hashes, paths, fitted thresholds, or final decimal places. Compare the structural facts described here rather than copying placeholders. If the structure differs, stop and inspect the earliest unexpected line.

1. the classifier receives two features in a fixed positional order.
2. max_depth=2 limits the number of learned questions along a path.
3. fit searches splits using only the first five rows.
4. the final row is held out; export_text exposes the learned thresholds instead of relying on a diagram.

These lines form one chain: numeric rows [hours, practice_tests] with readiness labels becomes class predictions, a readable tree, and held-out accuracy. Change only one input first. When several values change together, you cannot tell which change caused the new behavior.

• First failure: reversing feature order at prediction time silently changes the question each column answers. Re-run the smallest command that proves the repair.
• Second failure: an unconstrained tree can create leaves containing one training row and memorize noise. Preserve the failing input as a test when it represents a realistic mistake.
• Misleading success: text categories in numeric columns require consistent encoding rather than arbitrary integer labels. A clean-looking final line cannot cancel contradictory intermediate evidence.

When debugging, copy the exact error text and inspect names, paths, shapes, types, and versions. Explain the cause in one sentence before changing code. That discipline prevents a guessed repair from creating a second defect.

The deliverable for this step is a depth-limited classification tree with held-out accuracy and printed decision rules.

For the current ship step, save the smallest useful evidence: the relevant command, its output, and the input that produced it. Do not use a screenshot as the only record when text can be copied and searched. Keep generated artifacts separate from source inputs so rerunning the example does not destroy the evidence it is meant to evaluate.

Readable rules are not automatically fair. Audit proxy features, missing-value handling, and outcome rates across relevant groups. Version the feature schema because positional drift may not raise an error.

Return to your opening prediction. Mark it correct or rewrite it with the condition you missed. Then explain the difference between a successful execution and a trustworthy result for this specific example.