A specialty chocolatier uses a vision model to inspect truffles on its packaging line before they go into gift boxes sold at a premium price. Discarding a perfectly good truffle that the model mistakenly flags as defective wastes an expensive, hand-finished piece, while a genuinely defective truffle that slips through is rare and is usually caught by a human at the next station anyway. Which evaluation metric should the team prioritize when tuning the model's decision threshold?
Select an answer to reveal the explanation.
Short Explanation
Start by asking which mistake actually costs the business money. Here, tossing out a flawless, hand-finished truffle because the model got spooked is the expensive error, since that piece is gone for good. A defect that slides through, on the other hand, gets caught by a person standing at the next station, so it barely matters. That asymmetry points straight at one metric: out of everything the system decided to reject, how much of it actually deserved to be rejected? That is what you want to push as high as possible, even if it means letting a few true defects slip past, because those are cheap to catch elsewhere. The tempting wrong turn is to instead chase catching every single defect no matter what, which sounds responsible but actually maximizes the costly kind of mistake in this situation. And don't reach for tools built for predicting a number on a continuous scale, like a dollar amount or a measurement; this is a sort-into-two-buckets problem, not a predict-a-number problem, so those numeric-error metrics simply don't apply here.
Full Explanation
The correct answer is B. Precision measures, out of everything the model flagged as defective, how many were actually defective; maximizing precision directly reduces the number of good truffles wrongly discarded, which is exactly the costly mistake described in the scenario. Option A is incorrect because prioritizing recall would push the model to catch every last defect at the expense of raising more false alarms, which in this scenario is the wrong tradeoff since false alarms are the expensive error and missed defects are cheaply caught downstream by a human. Option C is incorrect because mean squared error applies to regression tasks that predict a continuous numeric value, such as a price or a temperature, not to a classification task that sorts truffles into defective or acceptable categories. Option D is incorrect for the same reason as option C: R-squared describes how well a regression model's predicted values track actual observed values, which has no meaning for a model producing a defective versus acceptable classification. Because this is a classification problem with an asymmetric cost that favors avoiding false positives, precision is the metric that should guide threshold tuning.