False reject rate: what the line feels, not the accuracy on the slide

False reject rate, also called overkill, is the proportion of conforming parts an inspection system rejects. Its twin is the escape: a non-conforming part it passes. Both are error rates, but only one lands on the floor that shift, and that asymmetry is why the false reject rate decides whether operators keep using the cell.

Headline accuracy hides it, because defect rates are small. If 1 part in 500 is genuinely non-conforming, a system that simply passed everything still scores 99.8%. Accuracy on an imbalanced population is a statement about how rare defects are, not about how well the cell judges them.

Two errors, two owners, two discovery times

The two rates move against each other: tightening sensitivity to shrink escapes grows overkill, and loosening it does the reverse. That is why the setting is a governed choice rather than a tuning exercise — see the anomaly score is a ranking, the threshold a business decision.

Convert the percentage into a pile

A percentage is not reviewable by the people who live with it. Convert it against real throughput. The figures below assume 1,200 parts per hour on an 8-hour shift — 9,600 parts — and 45 seconds to retrieve, re-judge and dispose of each pulled part.

Three things become arguable once the table exists: whether the re-check is staffed, whether the rate is acceptable per class or only on average, and whether anyone will notice the gap between 0.5% and 2% before the trial ends. Systems trained on conforming parts alone start high on overkill and improve as real rejects accumulate — the migration in when you have forty defect images, and when you have four thousand.

The re-inspection loop, and who staffs it

1. The part leaves the line to a destination that is neither the scrap bin nor a tray on the floor, or the evidence is gone before anyone can use it.
2. Someone judges it against the class list, with the golden and limit samples in reach — a named role on the shift plan, not whoever is free.
3. The verdict is recorded against the part's image and score, agreed or overturned, with the class. This is the only record a true overkill rate can be computed from.
4. The disposition is executed, and the queue depth is watched: a queue that grows all shift means the rate is above what the station can absorb.

If that loop is unstaffed the failure is quiet and predictable: pulled parts accumulate, someone starts returning them unchecked, and within a fortnight the cell is worked around rather than used. The station also needs somewhere to log a verdict in seconds — ordinary internal tools and ops work, nothing to do with the model.

A system is not abandoned in a meeting. It is abandoned by an operator who has learned that the reject bin is mostly good parts.

The number to agree before the build starts

Write the overkill budget in the units above and per class, not as one percentage: the acceptable pile for a class dispositioned scrap differs from one dispositioned use-as-is. Agree who staffs the re-check at that volume, and what happens if the rate lands at twice the budget in week 1. Trials that skip this run months on one line with still no decision, because nobody wrote down what good looked like in shift units.

The rest of the imaging, threshold and acceptance decisions sit under visual inspection and defect detection, and the on-premise vision work under manufacturing and industrial vision.

• overkill
• inspection metrics
• line operations
• quality cost