Automated inspection is often pitched as a way to take people out of quality control. In practice, the people on the line know things no model has seen yet: a new supplier's surface finish, a fixture that was adjusted this morning, a defect type that has only appeared twice. An inspection system that ignores that knowledge stays exactly as good as the day it was deployed.
Entroptics Vision is built on the opposite assumption: the machine makes a first call, and people decide what happens when the machine is unsure.
Route uncertainty to people
Not every result needs a human. Clear passes and clear defects can flow straight through. Uncertain findings are different, and Vision keeps them visible for review at the station, with the image, the machine verdict, and the supporting evidence side by side.
Reviewers can confirm a finding, correct it, or request another look. The goal is a clear next step for the operator, not another alarm to dismiss.
Never overwrite the machine's answer
When an operator corrects a result, Vision records the correction separately from the original machine verdict. Both are kept. That matters for three reasons:
- Auditability. You can always see what the machine said, what the person decided, and when.
- Measurement. Comparing machine and human judgments over time shows where a model is drifting and where it is reliable.
- Learning. Corrections become labelled examples, without losing the record of what the model originally got wrong.
Traceability means every result links back to its image, station, part, revision, and review history. You can follow how a finding became a quality decision.
Turning corrections into better inspection
Operator feedback is only valuable if it reaches the model, and only safe if it reaches the model under control. In Vision, approved annotations feed versioned datasets. A newly trained model is evaluated against a separate held-out dataset before it can be promoted, and an approved artifact is packaged before it is assigned to stations.
The same principle applies beyond models. Confirmed root-cause investigations can support proposed changes to diagnostic rules, and two distinct people sign off before those changes become part of the knowledge base.
Better decisions, not fewer people
The measure of an inspection system is not how many decisions it takes away from people. It is how much better the decisions become over time. Keeping operators in the loop, recording what they know, and feeding it back under review is how an inspection keeps up with a line that never stops changing.