Updated July 2026 · 7 min read · Adente Vision Engineering Team
Why acceptance-test on your own parts, not the vendor's demo?
A vendor demo proves the system works on the vendor's parts under the vendor's lighting, which is not the question you need answered. Your parts, your line speed, your lighting and your defect mix are what the inspector has to survive in production, and the only defensible way to know it will is to measure it on those parts before sign-off. An acceptance test turns "it looked good in the demo" into a signed number.
The structure is the same as any measurement-system acceptance: define what pass means before you start, run a controlled set of parts through the system, and check the results against the criteria you agreed. Treat the inspector as a gauge you are qualifying, not a gadget you are trying out. 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, and the sections below lay out an acceptance test you can run with your integrator on your own line.
How do you design the pilot part set?
The pilot part set has three groups, and you assemble all three before the run so nobody tunes the system to the test afterward. Each group answers a different question, and each maps to a metric you will record.
- Known-good parts: conforming parts that the system must pass. These measure the false-reject rate, the good parts wrongly rejected.
- Known-bad parts: genuine defective parts pulled from your scrap history. These measure the false-negative rate, the defects that escape.
- Seeded-defect parts: parts with deliberately introduced defects of each type you care about, at the severity threshold you need caught. These measure detection rate by defect class and probe the borderline.
Size each group so the numbers mean something. A handful of parts cannot establish a rate; agree the counts with your quality team so the escape and false-reject figures have enough parts behind them to be credible. Because the model trains on good parts only from about 20 good reference images, keep the training images separate from the acceptance parts, so you are testing what the model generalises to, not what it memorised.
Which metrics decide pass or fail?
Two numbers decide the accuracy question: the false-negative rate, the share of real defects that escaped, and the false-reject rate, the share of good parts wrongly rejected. F1-score combines the underlying precision and recall into one figure, but on an acceptance test you want the two error rates called out separately, because they carry different costs and different owners on your line.
Record them on your parts, not the vendor's. As a reference for what strong looks like, Adente Vision reached 99.65% F1 with a 0.69% false-negative rate on a live cap-inspection line, but that is the vendor's context; your acceptance number is whatever the system scores on your seeded and known-bad sets. The stage table below ties each pilot group to the metric it produces and the criterion it must clear.
| Pilot stage | What you measure | Pass criterion |
|---|---|---|
| Known-good run | False-reject rate (precision) | At or below your agreed false-reject ceiling |
| Known-bad run | False-negative rate (recall) | Zero escapes on the critical defect set, near a low FNR reference |
| Seeded-defect run | Detection rate by defect type | Every seeded class caught at the set confidence |
| Repeatability (gauge R&R) | Measurement variation, same parts repeated | Within the standard method's acceptance guidance |
| Speed under takt | Per-part inspection time | Inside your cycle budget, measured on your line |
How does gauge R&R apply to a vision check?
Gauge R&R (repeatability and reproducibility) asks whether the same part, measured repeatedly, gives the same result, and it applies to a vision inspector as much as to a caliper. Run a set of parts through the system several times, and across shifts or operators where relevant, and record how much the pass/fail or the measured dimension varies when nothing about the part has changed. Variation with no real cause is measurement noise you need to bound before you trust the gauge.
Use the standard method rather than an ad-hoc check. The AIAG Measurement Systems Analysis reference gives the accepted procedure and the common acceptance bands, where a study result under about 10% is generally acceptable and 10% to 30% is marginal depending on the application. For dimensional checks, the related rule-of-ten (or rule-of-tens) guidance holds that the measurement system should resolve to roughly a tenth of the tolerance it is judging, so decide up front that your inspector's resolution is fit for the tolerance on the drawing. These are standard-method figures from measurement literature, not Adente claims, and they give your acceptance test an external basis instead of an in-house opinion.
How do you validate what a 20-image model actually learned?
Training on about 20 good images is efficient, and it makes validation more important, not less, because you want to confirm the model learned the defect boundary rather than a quirk of the training set. The seeded-defect run is where you check this: introduce defects the model never saw in training, at and just above the severity you need caught, and confirm it flags them at the confidence threshold you set before the run.
Set that confidence threshold first, then run, so you are not moving the goalposts to make the system pass. Feed the model any borderline or missed cases from the seeded run as new good or bad references, retrain within the under-48-hour window, and re-run the acceptance set to confirm the fix held without breaking the known-good rate. When every stage in the table clears its criterion on your parts, the sign-off is defensible. For the metrics behind the pass criteria, see the sibling post on the false-negative rate; for how to make a vendor state these numbers up front, see the vendor-benchmark post; and for the method behind few-shot and anomaly detection, see the pillar guide.