Why manual visual inspection drifts — and what to measure
By HoldField · 2026-06-18
Manual visual inspection is not unreliable because inspectors are careless. It drifts because it depends on attention, lighting, fatigue, and an internal sense of the threshold that no two people hold identically — and that no one person holds identically across a shift. Treating that variation as a training problem alone misses most of it.
Where the variation actually comes from
- Between inspectors: two people apply the same written criterion differently at the borderline.
- Within an inspector: the same person judges differently at hour one and hour seven.
- Environmental: a change in lighting or fixture position shifts what a defect looks like.
- Rate pressure: as line speed rises, dwell time per part falls and marginal defects are missed first.
The measurements worth having
You do not need a large study to see drift. Two lightweight measurements surface most of it. The first is agreement: periodically have two inspectors judge the same set of parts and look at how often they agree, especially on the borderline cases. The first is not about who is right — disagreement itself is the signal. The second is repeatability: slip the same known part back into the flow occasionally and check that it is judged the same way each time.
If a known-good and a known-marginal part are not consistently judged the same way through a shift, the process is telling you the threshold is moving — before any customer does.
What to do with the signal
Drift is not a reason to replace people; it is a reason to give the borderline a firmer definition and a consistent second look. The parts that vary most between inspectors are exactly the ones worth capturing as reference examples, because they mark where the written criterion is under-specified. Automation earns its place here precisely where human judgement is least repeatable — the borderline — while humans keep the calls that need context.