Sampling inspection: what a handful of parts can and cannot tell you about a lot

The two risks the scheme is balancing

Any sampling arrangement can go wrong in opposite directions. It can reject a lot that was actually fine, which costs the producer sorting, delay and argument. It can accept a lot that was not, which costs the buyer whatever the defect does downstream. Every scheme is a chosen point between those two, and no amount of sample size removes them both. Being explicit about which error hurts more for this part is the real decision. A low-consequence cosmetic feature can tolerate a loose arrangement, while a characteristic that stops an assembly line justifies a scheme that rejects far more readily.

Acceptance is a statement about the lot, only if the lot is real

The arithmetic assumes the sample is drawn from a homogeneous population produced under the same conditions. That assumption fails constantly in practice. A pallet built from several shifts, several cavities, two tool changes or two material batches is not one lot however the paperwork describes it, and a sample from it tells you about the pieces you happened to pull rather than about the whole. Where variation between subgroups is expected, sample against those subgroups. The commonest way sampling fails is not a mathematical error but a delivery assembled from whatever was on the floor.

Attributes and variables buy different amounts of information

Judging parts as pass or fail is simple, needs no gauge beyond a limit, and works for anything that can be classified. Recording actual measured values from the same parts is more effort and yields far more, because you see where the process is sitting relative to the tolerance and how much it is spreading, which lets you act before anything fails. For the same information about a lot, measured data needs fewer parts than pass-fail counting. Where the characteristic can be measured at reasonable cost, taking numbers rather than verdicts is usually the better purchase.

Where sampling has no business being used

Some characteristics do not tolerate a probabilistic answer. Safety-related features, anything where a single escape has consequences out of proportion to the lot value, and characteristics where defects arise from a sporadic cause rather than the process spread all fall outside what sampling can sensibly cover. A wrong-part mix-up is the clearest example: it is not distributed through the lot in any statistical sense, so a sample either happens to include one or does not. Detection for that class of problem has to come from control at source, identity checks or full verification, not from a scheme.

Published schemes, and the honesty they require

Standard sampling schemes are published by national and international standards bodies and give tabulated plans for a chosen quality level and lot size, which saves each plant deriving its own. They work when applied as intended, and they mislead when the quality level is chosen by copying a previous contract, when lots are assembled artificially, or when a rejected lot is resubmitted after a partial sort until it passes. That last practice destroys the guarantee the scheme offers, and it is common enough that many buyers write rules about resubmission into the purchase agreement.

Frequently asked questions

Does an accepted lot mean the lot contains no defects?

No. Acceptance means the sample gave no reason to reject, at a defined level of risk agreed in advance. A lot with a small proportion of defective pieces will pass a typical scheme most of the time, which is exactly what the arrangement was designed to allow. Customers who expect zero escapes from a sampling regime have misunderstood what they agreed to, and the discussion is best had at contract stage rather than after a complaint.

Why does sample size not scale with lot size the way people expect?

The information a sample carries depends mainly on the sample itself rather than on the size of the population behind it, so doubling a lot does not require doubling the sample. Published schemes do increase samples somewhat for larger lots, but far less than proportionally. This surprises buyers who assume a large delivery deserves proportionally more checking. The stronger lever is the acceptance level chosen, not the count of parts pulled.

Can we sample a delivery that arrived as one pallet from several production runs?

Not meaningfully as a single lot. Mixing runs breaks the assumption the scheme rests on, because variation between runs can be larger than variation within them and a sample may miss an entire bad run. Either require the supplier to keep runs separated and identified on delivery, or sample each identifiable subgroup. Where the supplier cannot say which parts came from which run, the traceability problem needs solving before the sampling question is worth discussing.