You already know: classification and calibration (D3), gradient-boosted trees (D3), explainability methods such as feature contributions (D10), and from the churn case that your decisions shape the data you will learn from. This case adds regulated, high-stakes decisions: outcomes take a year to observe, you only see outcomes for the people you approved, every denial must be explained, and the system must be fair by law. The design questions: how to define and obtain labels, which model families survive validation, how a probability becomes a decision, and how to change a governed model safely.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud. Legal requirements differ by country; this lesson describes common patterns, not legal advice.
1. The prompt, and the questions that change the design
A mid-level answer: "train a model to predict default and approve applicants below a threshold". A staff candidate first asks about the product and its economics (loan size, margin, what is lost when a borrower defaults), decision speed, regulation (explanations, fair lending, model governance) and labels (what counts as default, how long it takes to observe). Those answers produce an economic cutoff, an instant path with a manual review band, an explainable, monotonic model with reason codes, and a label plan for delay and selection bias.
Assumptions you state: