In this deep dive
A batch is not good because someone looked at it. It is good because the defect classes, the sample size and the decision rule were agreed before the first piece was made — and because the next batch can be compared with this one on the same numbers.
Principle
Checking is not a system
“We check every piece” is a statement about effort. It is not a statement about method, and the two are not the same thing. A hundred percent inspection has its own failure modes: an inspector working through a tray of five thousand pieces stops seeing the fifth defect, applies the standard differently on Thursday than on Monday, and leaves behind no record that anyone can argue with later.
Sampling has different failure modes, and they are at least visible. It accepts a known risk, in exchange for a rule that two people can apply the same way: a defined sample size, defects sorted into classes, an acceptance number, and a record.
So the useful question is not “how much do you inspect”. It is: what would let a buyer accept a lot they have never seen? The answer has to be a system, because a promise cannot be checked and a system can.
What this page is for
The five earlier guides describe how a part is made and what can go wrong at each stage. This one describes how a batch is accepted: what is measured, how many pieces, by which rule, and what happens when the rule says no.
Tolerance
What a tolerance can promise
A tolerance is not a wish. It is the space a process needs in order to be repeatable, and that space is created by the process itself — long before anyone writes a number on a drawing.
Where the space comes from: a die wears, so dimensions drift as the tool runs; plating adds to every surface it touches, so a plated dimension is not the machined dimension; casting needs draft and shrinks as it cools; hard enamel is ground back after filling, which removes material from the face. Each of those is a physical allowance, and none of them can be argued away by tightening a number.
This is why the same drawing gets a different tolerance from a stamped part and a cast part. It is also why a drawing should say which dimensions are functional — a hole position, a stem diameter, an overall size — and which are cosmetic. When every dimension carries the same tight tolerance, the inspection bench ends up enforcing a standard the process was never able to meet, and the argument moves from the drawing to the loading dock.
Classification
Defects have three weights
Before any sampling plan means anything, the defects have to be sorted. A scratch on the back and a clutch that lets go are not the same event, and counting them in one column makes the acceptance number meaningless.
Critical
Unsafe or unusable. A sharp edge, an attachment that lets go, a restricted substance above the limit.
Zero accepted, whatever the sample says.
Major
The piece cannot be sold as intended: wrong colour, a missing enamel cell, plating broken through, the wrong hardware.
Judged against the agreed AQL for the class.
Minor
Visible but acceptable at arm's length: a faint mark on the reverse, a slight polish trace on a rim.
Judged against a looser AQL, or accepted by count.
The classification has to be agreed before the inspection, not after a dispute. It is the only way two inspectors, in two places, reach the same count — and the only way a rejection can be explained without appealing to taste. If a supplier and a buyer cannot agree on which class a defect belongs to, they are not going to agree on the verdict either.
Sampling
AQL: from sample to decision
AQL is the most quoted and least understood number in this trade. It is not “the percentage of defects we allow”. It is the quality level at which a lot still has a defined, low probability of being rejected — a rule for turning a sample into a decision, with the risk stated rather than hidden.
Lot size
How many pieces are in the shipment. This is the only input that comes from the order.
Sample code
The standard's table converts lot size into a code letter, and the code into a sample size.
Draw the sample
Random, from across the lot — not the top layer, not the pieces the line set aside.
Accept or reject
Count the defects by class and compare with the acceptance number. One rule, applied to every lot.
Two risks sit inside that rule. A good lot can still be rejected — that is the producer's risk — and a bad lot can still be accepted, which is the buyer's. A tighter AQL reduces the second and increases the first, and it also increases the sample size. There is no setting where both risks disappear; there is only a setting where both are known.
Which is why the level has to be named in the order. The standard, the inspection level and the AQL for each defect class are agreed per order — usually the buyer's specification, or the one written into the quotation. When nobody names them, the inspection has no decision rule, and what is left is an opinion with a clipboard.
What AQL cannot tell you
It says nothing about the next lot, and nothing about whether the process is stable. It answers one question only: given this sample, should this lot be accepted? The next section is about the question it leaves open.
Inspection
The four inspection points
Inspection is not one event at the end. It is four events, and each one catches something the others cannot.
Incoming material
Base metal, enamel batch, plating chemistry. Nothing later in the process can fix a bad input — it can only hide it for a while.
First article
The first piece off the tool or the first production piece, measured against the approved sample and the drawing. Catches a setup error before it becomes a thousand pieces.
In-process
At fixed intervals during the run. Catches drift: tool wear, bath depletion, a colour that has moved since the morning.
Pre-shipment
The AQL gate. It catches what the earlier points missed, and it produces the record the buyer actually receives.
The economics of the four points are not equal. A fault caught at incoming material costs a phone call; the same fault caught at pre-shipment costs a sorted lot, a delayed container, and an argument about who pays for the sort. That is the entire reason to inspect early, and it has nothing to do with distrust.
Evidence
What a useful report contains
An inspection that leaves no record is a conversation. A record that cannot be tied to a carton is a story. What a buyer needs is a document that answers “which pieces, measured how, against what, and what did you find”.
Two numbers that tie the report to a physical carton.
Which sampling standard, which inspection level, which AQL per defect class.
How many pieces were drawn, and from where in the lot.
Critical, major and minor, counted separately — never as one total.
The functional dimensions, with the specification beside each reading.
The defects that were found, not only the ones that were not.
Accept, reject, or accept after sorting — stated plainly, with a signature and a date.
The two fields people leave out are the two that matter most in an argument: the sample size, and photographs of the defects that were found. A report that only shows the good pieces proves nothing except that the photographer chose well.
Drift
Consistency is not conformance
AQL compares a lot with a standard. Consistency compares lots with each other. Those are different questions, and a supplier can pass the first one for a very long time while failing the second.
This is what drift looks like from inside a quality system. A die wears, a bath depletes, a pigment lot changes, an operator compensates a little. Every individual lot can be inside the band while the process walks toward one edge of it. Then one shipment crosses the line, and to everyone on the buyer's side it arrives without warning.
Seeing it needs a different instrument: pick two or three functional characteristics, measure the same one on every lot, and plot the readings against the band over time. The trend appears long before the failure. That is also the answer to the most common complaint in this trade — “the second batch doesn't match the first”. It usually matched the standard exactly.
Why colour and plating drift first
They are the two characteristics with the most inputs: pigment batch, bath chemistry, plating thickness, polishing time. They are also the two the buyer notices first, which is why a consistency chart is worth more than another inspection.
Response
When a batch fails
A rejection is not the end of the system — it is the moment the system either works or doesn't. Four steps, in this order, and the order matters more than the speed.
Contain
Stop the lot. Mark it, move it, and keep it away from anything that could ship. The first hour decides whether the problem stays small.
Establish the extent
A 100% sort, or a larger sample, to find out whether this is one carton or the whole run.
Rework or remake
Re-plating and re-enamelling are cheaper than a new run, but they rarely restore the approved surface. Say which one you are doing, and why.
Record it
What failed, how many, what was done, who paid. Without the record, the same failure arrives again with the next order.
The part people skip is the last one. A failure that is fixed but not recorded is a failure that will be repeated, usually on the next order, usually with a different person explaining it.