No Number Changes by Accident

Problem. Factories review defect rates, scrap status, Pareto charts and alerts every week, and those figures often come from Excel/VBA tools written years ago by people who have left. Nobody knows exactly what they compute. When a team modernizes one, either the rewrite copies the old tool blindly, keeping errors nobody knew about, or it follows the written procedure and silently changes numbers people have trusted for years, and nobody can explain why. Both make modernization slow and erode trust in the new tool. Solution. A repeatable workflow, run with IBM Bob, that makes every difference explicit before any code changes: 1. Freeze the legacy outputs as a golden master. 2. Bob analyses the legacy code and lists where it differs from the official procedure. 3. Bob implements the procedure independently and compares every value with the legacy outputs, so differences are proven with numbers. 4. The product owner decides, for each difference, to keep the legacy behaviour or follow the rule. 5. Bob ports the tool with tests that require identical outputs. 6. Each decision becomes a switch; corrected mode applies them. 7. An app shows both modes and explains every changed number. Who it's for and how they use it. Developers and IT teams who must replace a legacy tool run the workflow with Bob in their repository, following the playbook and prompts in docs/WORKFLOW.md. The quality engineer fills in the decision log: keep or fix, for each difference. The quality team then uses the app: they switch between Legacy and Corrected, see the reports they already know, and open "What changed" to see why each number moved, before go-live. Why it's different. Most AI modernization demos rewrite code and hope. Here the old tool's own outputs are the test oracle, an independent implementation of the procedure is a second reference, and a person decides every difference. Bob's discovery was measured against a hidden answer key: 3 of 6 planted divergences found by reading the code, 5 by the rules check, 6 through the tests. Impact on the NC Tracker (a synthetic Excel VBA tool, 2009–2013, with quality rules R1–R7): - 9 divergences, each tied to a VBA line and proven with data: one line's defects silently halved since 2011, a 3.00 % scrap rate flagged RED instead of ORANGE, accepted parts missing from monthly figures, Pareto weeks starting on Sunday. - All five legacy reports reproduced exactly, so the switch-over is provably safe. - 358 changed values, each traced to its decision; 90 automated tests; tagged milestones from v0-legacy to v3-app. No number changes without a documented decision.