Instructor guide

Begin by asking learners to name a prediction time and demonstrate a wrong but available-looking outcome field. The first success criterion is an information boundary, not a fitted model. Use the readiness checks in the syllabus to redirect missing Python, SQL-grain or Statistics skills without restarting those courses.

Require learners to predict the result of each miniature trace before running code. Ask them to reconcile a confusion matrix to its denominator and explain an embargo with actual timestamps. A learner who can run GridSearchCV but cannot identify the rows that influenced a choice needs targeted review before the capstone.

Preserve the answer boundary: supply index.html and the raw package, but ask learners to postpone solutions.html, final-test outputs and the full invariant suite until the corresponding reveal. All files remain accessible; this is an instructional protocol rather than technical exam security. Do not represent reference smoke tests as an assessment of the learner's own work.

The course deliberately contains a mixed calibration result and a future-capacity failure. Ask what each means before accepting a deployment recommendation. The best operational decision can be to hold. The synthetic data is modestly imbalanced, has correlated inputs and a changed future process, and includes post-outcome leakage traps. It does not establish fairness, causality, production accuracy or the costs of a real intervention.

Use the 100-point assessment rubric and its critical safety gates. Proposed timings and thresholds require a real learner pilot. Record confusing passages, time to first successful run, installation failures, explanations of leakage and threshold selection, and whether learners can transfer the method to an altered prediction question. Collect feedback with appropriate consent before revising completion claims.

Optional extensions should remain bounded: vary the documented cost ratio using validation only; compare a changed predeclared training window with a new untouched future window; or design, on paper, an evaluation for unseen depots in a later period. Do not use the included final-test results as the target for repeated improvement. Clustering, high-stakes personal scoring and cloud deployment are outside this package.

For editorial rebuilding, install requirements-build.txt into an editing environment and run python tools/build_guide.py. Run python tools/build_visuals.py only after reference development results exist. Both generated HTML pages derive from the Markdown sources. Diagram generation uses the existing course dependencies; it does not require Matplotlib.