Published
A defect detector answers one question: is this unit acceptable? That is a useful question, and a system that answers it reliably pays for itself by keeping bad parts from reaching a customer. But it is a question about output, and the answer is discarded once the part is sorted.
A measurement system answers a different question: what, quantitatively, is this unit? The pass/fail decision then falls out of a threshold applied to the measurement — and the measurement itself survives the decision.
The difference shows up later
On day one the two look similar. Both reject the same parts. The divergence appears when someone asks why the reject rate moved.
A detector can say that it moved. A measurement record can often say what moved with it — that a dimension drifted in one direction across a shift, that the drift tracks a tool change, that it began before the rejects did. The first is a reason to investigate; the second is most of an investigation.
It changes what a borderline case costs
Classifiers are least reliable near the decision boundary, which is exactly where the parts that matter sit. A binary output gives no way to distinguish a unit that barely passed from one that passed comfortably, so that information is lost at the point it would have been most useful.
Retaining the underlying quantity keeps it. It also makes threshold changes cheap: tightening a tolerance becomes a configuration decision rather than a retraining exercise.
The cost of the framing
Measurement is more demanding. It needs calibration, it needs stable illumination, and it needs the measured quantity to be defined precisely enough that two engineers would agree on it. Detection can often be trained from labelled examples without any of that.
That cost is real and it is why detection is the more common framing. The argument is not that measurement is always right — it is that the choice should be made deliberately, because it determines whether the system is still generating value once the obvious defects have been designed out.
A reasonable default
Where the quantity of interest can be defined and calibrated, measure it and threshold the measurement. Where it genuinely cannot — surface appearance, texture, classes of defect that resist definition — detection is the honest choice, and a detector that reports a confidence is better than one that reports only a verdict.
The failure mode worth avoiding is reaching for detection because it is easier to get started with, and discovering later that the system cannot answer the question the plant actually has.