Updated July 2026 · 7 min read · Adente Vision Engineering Team
What are escapes and false rejects on a production line?
Every inspection decision falls into one of four boxes, and two of them cost money. A true negative passes a good part and a true positive rejects a bad one; both are correct. The two errors are the false negative, where a defective part is passed and escapes down the line, and the false positive, where a good part is rejected for no reason.
On the floor these are called the escape and the false reject. The escape is the defect that got away, on its way to assembly or the customer. The false reject, also called over-rejection, is the good part you scrapped or re-inspected for nothing. Every inspector, human or AI, produces some of both, and the mix between them is what sets the true cost of inspection, not the headline accuracy number.
Why does one threshold control both errors?
A single decision threshold turns a continuous score into a pass or fail, and moving it trades one error for the other. An AI inspector does not output "good" or "bad" directly; it outputs a score, how far a part deviates from normal or how confident it is that a defect is present. A threshold on that score makes the call.
Set the threshold strict, flag the faintest deviation, and you catch more real defects but also reject more good parts that happened to look unusual: fewer escapes, more false rejects. Loosen it and the reverse happens: fewer good parts scrapped, more defects slipping through. You cannot push both errors to zero with one knob. Every threshold is a chosen balance between escapes and false rejects, which is why "how accurate is it" is the wrong question and "where is the operating point" is the right one.
Why is an escape more expensive than a false reject, usually?
An escape and a false reject rarely cost the same, and on most lines the escape costs far more. A false reject is paid once, inside your plant: the price of one good part scrapped, or the labour of re-inspecting it. An escape is paid later and larger, because a defect that reaches assembly or the customer drags in rework, returns, warranty claims and, in the worst case, a recall.
That asymmetry is why "catch everything" feels safe, but a threshold tight enough to stop every last defect can reject good parts in numbers that quietly erode yield. The right question is not which error to allow, but what each one costs on your specific part, so you can weigh them against each other instead of guessing. A high-value safety part and a cheap high-volume part sit at opposite ends of that trade-off. The prevention, appraisal and failure costs behind both are an industry framing, not an Adente figure. For the escape metric itself, see the inspection metric vendors don't advertise.
Who pays for each error, and where does the threshold move it?
The four outcomes plus the two threshold moves fit in one view. The top four rows are the confusion matrix in shop-floor terms: who pays for each outcome and what drives that cost. The bottom two rows are the same threshold acting dynamically, showing how one knob shifts the mix.
| Error or move | Shop-floor name or effect | Who pays | Cost driver |
|---|---|---|---|
| False negative | Escape | The customer, then you | Warranty, returns, recall, lost trust |
| False positive | False reject | Your plant | Scrap, rework, re-inspection, yield loss |
| True positive | Caught defect | Your plant, cheaply | One contained reject, handled on the line |
| True negative | Correct pass | No one | A good part shipped |
| Tighten the threshold | Fewer escapes, more false rejects | Your plant | Yield falls |
| Loosen the threshold | More escapes, fewer false rejects | The customer, then you | Escapes rise |
Read the bottom two rows together and the trade-off is plain: because a single decision threshold sets both rates, chasing every last escape spikes false rejects and yield falls, while chasing zero false rejects lets escapes climb toward the customer. You choose an operating point, not zero of both.
Why does a low false-negative rate matter more than headline accuracy?
The false-negative rate measures escapes directly, which is why it beats accuracy as a trust metric on a rare-defect line. Accuracy blends every correct call together, so a line that is 99% good parts scores about 99% accuracy even if it misses most defects. The false-negative rate ignores the easy good parts and asks one thing: of the defects that were actually there, how many got past?
On a live cap-inspection line, Adente Vision holds a 0.69% false-negative rate at 99.65% F1, catching broken, unclosed and hinge-damaged caps in about 30 ms per part. The 0.69% is the number a quality manager can act on, roughly seven escapes per thousand defects, while the 99.65% F1 confirms the low escape rate was not bought with a flood of false rejects. Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators.
How do you set the operating point to your cost of quality?
Setting the threshold is a business decision before it is a technical one: put a cost on each error, then place the operating point where their combined cost is lowest. Name who pays for each error on your line and what drives that cost. An automotive or medical part, where an escape can trigger a recall, argues for a strict threshold and a very low false-negative rate, accepting some extra false rejects as cheap insurance. A low-value, high-volume part where scrap is expensive and an escape is a minor return argues for a looser point.
A better inspector does not just slide along this trade-off; it shifts the whole curve, catching more defects at the same false-reject level, which is what a low false-negative rate at a high F1 represents. The same threshold logic applies to every task the unit runs, defect, anomaly, counting and quality, across the inspection applications it covers. To put a currency figure on a single escape, see the companion post on the cost of a missed defect; for the wider method, read the pillar guide on AI visual inspection.