Legacy modernization becomes dangerous when a team changes code before it can explain the behaviour already crossing a business boundary. This focused, practical course uses Claude Code and a synthetic Python settlement lab to show a safer sequence. You will begin with a one-penny difference between a legacy batch and its proposed replacement. Rather than hiding the mismatch or asking an AI coding tool to rewrite the application, you will reproduce it, bound the workspace, and record a trustworthy starting point.
Next, you will trace the rounding behaviour, separate observed behaviour from business authority, and generate a scoped characterization record. You will then authorize one narrow correction in a reviewed worktree. The legacy implementation and golden master remain protected while you inspect the complete Git diff and run the differential gate. Finally, you will assemble a small evidence bundle and rehearse an exact patch-based rollback. The goal is not more documentation. The goal is a bounded claim another developer can inspect and reproduce.
The companion Northstar lab uses fictional records, Python 3.11 or later, GNU Make, and Git. It has no third-party Python dependencies and makes no network calls during the core exercise. This pilot is for developers, technical leads, and architects responsible for older systems with weak automated tests. You should already be comfortable with basic programming and Git. Claude Code is required for the guided analysis and bounded implementation session.
