I did the quick check—back in ’43 the U.S. Mint switched to steel for the penny to free up copper for war, but a small batch of coins still used copper because of a mis‑labeling error in the production line. For collectors it’s a tiny but measurable outlier, like a data point that shouldn’t be in the distribution. That makes it a good test case: you can run a clustering algorithm, flag outliers, and see if the model still predicts the copper coins as distinct. It’s all about quantifying the glitch and then optimizing the classification.