Manufacturing Quality Control Guide

Quality is not inspection. Inspection measures whether the system worked. The controls that decide the outcome are the specification, the approved reference sample, the materials approval and the pre-production review — and every one of them happens before there is anything to inspect.

That ordering is the whole subject. A buyer who books a pre-shipment inspection and calls it quality control has arranged to be told, at the last possible moment, what a decision made weeks earlier produced. By then the quantity exists, the materials are consumed, the vessel is booked and every remaining option is expensive. Measurements do not change the thing they measure.

The four controls that decide the outcome

1. The specification

   A written definition of the product precise enough that two people in different countries reach the same verdict on the same unit. Every ambiguity in it is a decision you have delegated to the factory, and it will be resolved in the direction of cost.

2. The approved reference sample

   The physical version of the specification, approved against a stated revision of it, held in identical copies by you, the factory and whoever inspects. It is what accept and reject actually mean.

3. Materials and component approval

   Approval of the bulk material and every component before cutting. Most defects that look like workmanship start as a substitution nobody approved.

4. The pre-production review

   A gate before the line runs: correct specification revision, approved sample present, materials in house and approved, inspection plan and defect classes agreed.

Inspection is the fifth thing, and it is the one that reports on the other four. Arranged in that order, most quality work stops being adversarial: you are not hunting for defects at the end, you are checking that a set of agreements is still being honoured.

Specification control: most quality disputes are specification disputes

When a shipment is rejected and the factory is genuinely surprised, the usual cause is not dishonesty. It is that both parties were working to different readings of the same document, and neither reading was unreasonable. "Heavyweight cotton", "match the swatch", "standard packing", "must pass QC" — each is an instruction that cannot be failed, because it cannot be tested.

The discipline is to write requirements that can be measured, by a stated method, against an agreed limit. A useful test before releasing a specification: for every requirement, name the instrument, the method and the pass condition. Where you cannot, you have found the clause the dispute will be about.

Tolerance

The permitted departure from a nominal value, stated together with the method and the point of measurement. A dimension without a tolerance is not a requirement; it is an aspiration that becomes an argument the first time somebody measures it differently.

The figures above are illustrative, to show the form of a testable requirement. Your own values come from your product, your market and what your customer will accept.

The approved reference sample, and its revision control

Approved reference sample

The physical unit both parties agree represents the specification — often called the golden sample. Sealed, dated, signed, and identified by the revision of the specification it was approved against.

A reference sample solves the problem that words cannot describe a hand feel, a sheen, a stitch density or an acceptable shade. It also introduces a problem words do not have: it is a physical object, so it fades, wears, gets borrowed, gets shipped to the wrong office and gets quietly superseded. Revision control is what keeps it a control rather than an heirloom. What a golden sample is covers the mechanics; the governance is short:

• Three identical copies, sealed and signed: one with you, one with the factory, one with whoever inspects. An inspector judging against a sample they have never held is guessing.
• Every copy labelled with the product, the specification revision, the approval date and who approved it.
• A superseded sample withdrawn or clearly marked, never left on the shelf beside the current one.
• A new revision is a new approval, with a new sample and a new date. Editing the old one destroys the evidence of what was agreed when.

The most dangerous state a sample can be in is "approved with comments". The physical reference and the written specification now disagree, and production will follow whichever is nearer the machine. Either incorporate the comments into a new sample, or write them into the specification and annotate the sample to say so.

Materials and component approval

Approving a swatch is not approving the bulk. Materials vary between lots, dye batches shift, a mill substitutes a yarn, a trim supplier changes a mould, and a component that is visually identical performs differently. The control is to approve what will actually be used, before it is cut, against the specification rather than against the previous delivery.

• Bulk approval from the actual production lot, not from the sample-room roll or the hanger.
• Every component on the bill of materials named with a supplier and a reference, including the ones nobody photographs — thread, interlining, labels, fasteners, adhesive.
• Test reports read for what they cover: an article, tested on a date, by a laboratory, against named criteria. That is a statement about the tested article, not about your shipment.
• A written rule for substitution: nothing is swapped without written approval, and a substitution offered late is a schedule problem being presented as a materials problem.

Where certification matters commercially, know which document proves what. GOTS, for example, separates the scope certificate — a site audited for named processes — from the transaction certificate, issued for the shipment itself. Components and trims then cause a disproportionate share of failures because they are cheap, sourced late and bought by whoever is nearest — exactly the condition for a substitution nobody logged.

The pre-production review

The cheapest gate in the system, and the one most often skipped because it feels like a meeting about things everybody already knows. Its purpose is to establish, while the quantity is still zero, that the factory is about to make the product you specified in the materials you approved under the rules you agreed.

• Before the line runs
• The specification revision in the factory is the current one, and everybody names the same number.
• The approved reference sample is on site, sealed, and matches that revision.
• Bulk materials and all components are in house and approved against the specification.
• Measurement points, methods and tolerances agreed, and the factory measures the way the inspector will.
• The defect classification agreed in writing and illustrated, and held by the factory’s own inspectors.
• The inspection plan agreed: which stages, who inspects, on what basis, and what each outcome triggers.
• Packing, labelling and market-specific requirements approved in physical form, not described.
• A first-off unit from the real line checked and signed before the run continues.

During production

In-line checks exist for one reason: while the line is running, the remaining quantity can still be different. A systematic error found at ten per cent complete is a correction; the same error found when everything is packed is a negotiation. The value of an in-line check is therefore almost entirely a function of how early it happens.

• A first-off check on the first units off the line, against the approved sample, before volume builds.
• The first pieces of each new size, colour, material lot and machine setting — changeovers are where systematic error enters.
• Measurement audits during the run, taken by the factory and spot-verified, not reconstructed at the end.
• What the factory is rejecting and why: a line producing no rejects at all is not inspecting.

Ask for in-line results as they happen rather than as a report afterwards. Data that arrives after the decision it was meant to inform is documentation, not control. The checks that make up a quality plan sets out the sequence in more detail.

Pre-shipment inspection

A pre-shipment inspection is carried out when production is substantially complete and packed, and it answers a narrow question well: does this finished lot conform to the agreed specification, at the agreed acceptance criteria, as presented? Two conditions make it worth what it costs. It has to happen while you still have leverage — before final payment and before the goods leave — and the criteria have to have been agreed beforehand. An inspection whose rules are settled afterwards is not an inspection; it is the opening position in a claim.

How acceptance sampling actually works

Acceptance sampling is the part of quality control most often described wrongly, usually as "the percentage of defective goods you agree to accept". It is a real and well-documented statistical method, and the shorthand misses the mechanism in a way that costs buyers money.

Acceptance quality limit (AQL)

A quality level that a sampling scheme is indexed to — expressed as an average, over a continuing series of lots from one supplier and one process. It is a parameter used to derive sample sizes and acceptance thresholds, not an allowance for defects inside any one shipment.

The standard is ISO 2859-1, "Sampling procedures for inspection by attributes — Part 1: Sampling schemes indexed by acceptance quality limit (AQL) for lot-by-lot inspection". Its third edition was published in January 2026, replacing the 1999 edition and its amendments, and it defines a system of acceptance sampling plans for inspection by attributes — single, double and multiple plans, with skip-lot procedures added in this edition. ISO describes the AQL as the level used "to determine sample sizes and acceptance/rejection thresholds", and the operative word in every plain-English gloss of it is average.

Sampling plan = f(lot size, inspection level, AQL) → sample size n, accept at ≤ Ac, do not accept at ≥ Re

Every term on the left is agreed before the inspector arrives. The standard supplies the tables that turn them into a sample size and the accept and reject numbers; this page deliberately does not reproduce them.

Read left to right, that is the entire mechanism. A lot is defined. Its size and an inspection level produce a sample size. The sample size and the AQL produce an acceptance number and a rejection number. Units are drawn at random, inspected against the agreed criteria, and the count of nonconforming units decides whether the lot is accepted. Nobody exercises judgement at the end, which is the point: it was exercised when the plan was chosen.

Two features of the scheme are what make the "acceptable percentage of defects" reading wrong. First, it is built for a continuing series of lots from a single process or supplier, not for one isolated shipment; a one-off lot is a different statistical question, and other parts of the ISO 2859 series exist for it. Second, the scheme contains switching rules that monitor results over time: consistently good quality earns reduced or skip-lot inspection, an increase in nonconformities forces tightened inspection, and continued poor quality leads to discontinuing acceptance sampling altogether. The AQL is the level a producer’s process average must stay at least as good as to keep having lots accepted — economic pressure applied through non-acceptance, with an upper limit on the buyer’s risk from the occasional poor lot.

The standard is also careful about what an AQL does not license: designating one gives a supplier no right to knowingly ship a nonconforming item, and an accepted lot does not oblige a buyer to accept a nonconforming unit found in it. Nor does non-acceptance mean destruction — the party responsible for the decision chooses the disposition, which may be sorting, rework, re-grading, holding or scrapping, and a corrected lot that is resubmitted is inspected again on stated terms rather than waved through.

Three things follow for a buyer. The AQL and the inspection level are contractual terms and belong in the order, not in a phone call on the day. Different classes of nonconformity are treated separately, with their own criteria — the most serious class attracting a much stricter one. And whoever specifies the plan is exercising an authority the standard expects to be named: your supplier’s own quality function, you, or an independent third party are not equivalent.

Measurement tolerances

A measurement dispute is almost never about the ruler. It is about where the measurement starts, how the unit was laid out, whether it was relaxed first, how many units were measured and which of them counted. A tolerance is only as good as the method attached to it.

• Name and illustrate every point of measure, so that "length" cannot mean two distances.
• State the method as well as the value: the instrument, how the unit is positioned, and any conditioning before measuring.
• Agree who measures, how many units, and what happens when the factory’s figures and the inspector’s disagree.
• Treat dimensional change after washing, heat or assembly as a separate control with its own test.

Defect classification, agreed in advance

The same mark is a minor defect to the factory and a major to the buyer, and both positions are sincere. Classification is therefore something you agree before production, in writing, with photographs — because after production the classification decides who pays, and nobody negotiates that neutrally while looking at the goods.

ISO 2859-1 handles this by classifying nonconformities by degree of seriousness and giving each class its own acceptance criteria, with a much stricter value for the class of highest concern, and it notes that the number of classes and what goes in them should suit the specific situation. It also draws a distinction worth borrowing: a nonconformity is the non-fulfilment of a specified requirement, while "defect" carries a definite meaning in law and is best avoided as a general term. In practice the industry says "defect", and the classification below is the one most buyers use.

Build the classification as a catalogue rather than three sentences: a photograph of the acceptable and the unacceptable case for each recurring defect in your product family, given to the factory’s own inspectors as well as yours. The boundary between major and minor is where the money is, so it is the boundary to illustrate most carefully.

Corrective action: how a failure becomes a documented change

A rejected lot is information. Whether it is also an improvement depends entirely on whether anything in the documents changed afterwards.

1. Contain

   Decide what happens to the affected goods and to anything else made under the same conditions — sort, rework, hold or scrap — and record who decided.

2. Find the cause

   Look for the mechanism, not the culprit. A defect that recurs in the same place is a process, a setting, a material or an instruction; blaming an operator guarantees the same lot again.

3. Change something durable

   The output of a real corrective action is an edit: to the specification, the reference sample, a work instruction, a jig, an incoming check or the inspection plan. A promise is not a change.

4. Verify and close

   Check the specific thing that failed, at the stage where it failed, on the next production, and record what changed and by what result. An open corrective action with no verification is an unresolved risk with a tidy reference number.

The test of the whole loop is simple and slightly uncomfortable: point at the document that is different because of the failure. If nothing is different, the failure will happen again.

Supplier performance data over time

One inspection tells you about one lot. A record across lots tells you about a supplier, and it is the only thing that turns quality control from a cost into a sourcing advantage. ISO 9001 requires an organisation to determine and apply criteria for the evaluation, selection, monitoring of performance and re-evaluation of its external providers, and to keep documented information about those activities — a formal way of saying that a supplier is qualified as of a date, and that the record is what re-qualifies them.

• The plan used per lot: lot definition, inspection level, and the criteria for each defect class.
• The result, and the disposition where a lot was not accepted.
• Nonconformities by class and by cause, so a recurring mechanism becomes visible across orders rather than within one.
• Measurement deviations by point of measure, which locate a process problem faster than a defect count does.
• Rework, replacement and expedited freight costs, and which party absorbed them.
• Corrective actions raised, what changed, and whether the following lot cleared.

That record buys three things. It lets you inspect proportionately — tighten on a supplier who is drifting, lighten on one consistently better than the agreed level, which is the same incentive structure the sampling standard builds in. It makes the next sourcing decision evidential rather than anecdotal. And it turns a negotiation about quality from two opinions into one dataset. Reducing supplier quality risk covers the supplier-management half, and factory vetting the qualification it re-opens.

A worked example, illustrative

A brand specifies a jacket with "waterproof zip, navy, match swatch". The factory sources a water-resistant zip that matches the swatch under its own lighting, and the pre-production sample is approved by email from a photograph. Bulk runs. At pre-shipment the zips are a visible shade off in daylight and fail the water test the brand assumed was implied. Nothing here was a workmanship failure: the specification named neither a component reference nor a test, the sample was approved from an image rather than an object, and the acceptance criteria were never written down. What prevents a recurrence is not a stricter inspection. It is a part number, a named test with a pass condition, a physical approval under a stated light source, and a catalogue entry showing an acceptable and an unacceptable shade.

Common mistakes

• Booking a pre-shipment inspection and calling it quality control.
• Approving a sample from a photograph, then rejecting a shipment for something a photograph cannot show.
• Agreeing the defect classification after the inspection instead of before production.
• Reading AQL as an allowance for defects in your shipment.
• Copying a sampling table from a blog and inspecting to a plan nobody can reconstruct.
• Approving materials from a swatch and never seeing the production lot.
• A tolerance with no measuring method attached to it.
• Closing a corrective action with a promise rather than a documented change.

The quality control checklist

• For one order, in order
• A specification whose every requirement names a method and a pass condition.
• Tolerances with measuring points and methods, graded across sizes where relevant.
• A defect classification agreed in writing, illustrated, and held by both parties’ inspectors.
• An approved reference sample, sealed and dated against a specification revision, in three identical copies.
• Bulk materials and every component approved against the specification before cutting.
• A pre-production review completed, with a signed first-off unit.
• An in-line check early enough that the remaining quantity can still change.
• The inspection plan agreed before inspection: stage, lot definition, inspection level, AQL per defect class, sample size and accept and reject numbers.
• Inspection timed while payment and shipment are still leverage.
• A disposition rule for a lot that is not accepted, and terms for resubmission.
• A performance record per lot, aggregated per supplier, that outlives whoever ran the order.

How Library of Trade approaches it

Our own sequence puts the controls before the inspection: a brief becomes specifications, tech packs and a bill of materials; materials are reviewed against them; samples are developed, measured, commented on and approved before anything is cut; and production is inspected against those agreed specifications and tolerances rather than a standard invented at the end. Depending on requirements we can support sample inspections, during-production checks, final inspections before shipment and quality reporting. AI structures and compares the information this generates; human experts make the calls that matter, including whether a shipment is approved. Brands contract directly with their suppliers throughout, and keep their specifications and product information.