A model I was working with scored R squared 0.998 on public US data. One of the input columns had been calculated from the target by the same agency that published both, so the model was rediscovering a formula it had already been handed. Tests don't catch this. Nothing throws, nothing fails, the numbers look great. So I built a linter for it. You write a manifest of what derives from what, the tool walks the graph, and CI goes red when a feature sits on a derivation path to or from your target. It returns four verdicts, including one for columns whose provenance nobody has traced yet. The code is Python. The idea applies anywhere you have a dependency graph and a CI job. I'll show the leak, the manifest, the linter, and a run going red live.
Repo: https://github.com/Adeniyikayodee/derives-from-tutorial