Using false accepts and false rejects correctly in engineering

A binary quality inspection has two fundamentally different error types. A false accept allows a defective part to pass. A false reject removes a good part. Both reduce process quality, but their consequences can be completely different.

False accept: the quality risk

A defect is classified as acceptable. Depending on the application, this can lead to rework at the customer's site, complaints, recalls or safety problems. The permitted threshold is therefore particularly low in critical applications.

False reject: the productivity risk

A good part is rejected incorrectly. The consequences are scrap, manual reinspection, lower line throughput and unnecessary cost. A system can therefore appear very safe and still be economically unusable.

Why accuracy is not enough

With imbalanced classes, accuracy may look excellent while the rare but important defect class is detected poorly. At a minimum, the confusion matrix, false-accept rate, false-reject rate, yield and the corresponding sample size should be considered together.

KPI limits should be derived from process risks, not retrofitted to the results after the test.

Four rules for defensible engineering decisions

1. State the consequences and costs of both error types.
2. Set and version limits before measurements begin.
3. Evaluate KPIs separately by scenario, variant and relevant disturbance factor.
4. Report confidence and data coverage alongside the percentage value.

From KPI to action

An exceeded limit is not a diagnosis. Only by connecting it to sensitivity analysis and worst-case samples can teams determine whether optics, illumination, data, parameters or decision logic should be changed.