Statistical QC in Luxury Handbag Production: Risk, Operating Characteristic Curves and the Numbers Behind an AQL Plan
What does statistical QC actually decide in a luxury handbag programme? It decides who carries the risk when only part of a lot is examined. At AQL 2.5 with a sample of 50 and an accept number of 3, a producer carries roughly a 5 percent chance of a good lot being rejected while a brand carries a far larger chance of a mediocre one being accepted. Understanding the operating characteristic curve behind those two numbers is what separates a sampling plan from a number quoted in an email.
The Decision Order, in Seven Steps
Statistical QC is a sequence of seven decisions, and making them in this order is what prevents the plan from being rebuilt during an argument about a shipment.
- Decide the question: are you testing a lot for acceptance, or a process for capability? They use different tools.
- Decide the unit: a piece is faulty or sound, or a piece is counted by the number of faults on it.
- Fix the inspection level, normally General Inspection Level II in ISO 2859-1.
- Read the sample size from the lot size and the level, and write the code letter down.
- Set the acceptance quality limit per defect class, with critical held at zero regardless.
- Agree the classification library before bulk, so the count is not negotiated afterwards.
- Agree the switching rules, so a rejection triggers a defined response rather than a conversation.
The first step is the one skipped most often. Acceptance sampling and process capability answer different questions, and a brand that uses an AQL plan to ask whether a process is stable will get a confident answer to a question it did not ask.
Glossary: The Twelve Terms That Actually Appear
Twelve terms cover almost everything that appears in a plan or a dispute, and each is worth knowing precisely rather than approximately.
| Term | What it means | Why it matters in practice |
|---|---|---|
| AQL | Acceptance quality limit: the quality level a sampling plan is built around | It is not a permitted defect rate, and reading it as one is the commonest error |
| Code letter | The letter derived from lot size and inspection level | It fixes the sample size, and quoting it makes a plan checkable |
| Ac / Re | Accept and reject numbers for the sample drawn | On 500 pieces at AQL 2.5, accept 3 and reject 4 |
| Producer risk | Chance a lot at the agreed quality is rejected, commonly near 5% | The risk carried by the production side in every plan |
| Consumer risk | Chance a lot at limiting quality is accepted, commonly near 10% | The risk carried by the brand, and usually the larger of the two |
| OC curve | Operating characteristic: probability of acceptance against true lot quality | It is the whole plan drawn as a line rather than stated as a number |
| AOQ | Average outgoing quality after rectification of rejected lots | Describes what actually ships, not what was submitted |
| AOQL | The worst average outgoing quality over all incoming levels | The honest ceiling on what a plan delivers over a season |
| Cpk | Process capability index against specification limits | Answers whether a process can hold a tolerance, which sampling cannot |
| Rectification | Rework and re-presentation of a rejected lot | It is what makes outgoing quality better than incoming quality |
| Switching rules | Rules moving inspection between normal, tightened and reduced | They make the plan respond to history rather than repeat itself |
| Limiting quality | The poor quality level accepted only with the stated consumer risk | The number a brand should ask for and almost never does |
Producer Risk and Consumer Risk
Every sampling plan distributes two errors, and the person who writes the plan decides how much of each error each side carries.
- Producer risk at the agreed quality level
- Commonly around 5%, meaning a sound lot is still rejected about one time in twenty
- Consumer risk at limiting quality
- Commonly around 10% at the poor-quality end of the curve
- Sample size on 500 pieces, Level II
- 50 pieces, code letter H
- Accept and reject at AQL 2.5
- Accept 3, reject 4
- Upper bound after zero defects in 50
- Roughly 5.8% at 95% confidence, which surprises most buyers
- Capability target where a tolerance is measured
- Cpk of 1.33 or better on a characteristic that can be measured
The fifth line is the one worth pausing on. A clean sample of fifty does not mean the lot is clean; at 95 percent confidence it means the true fault rate is below roughly 5.8 percent, which is a very different statement.
Reading the Operating Characteristic Curve
An operating characteristic curve is the plan itself: it plots the probability that a lot is accepted against the true proportion of faulty pieces in that lot, and every other number in the specification is a point on it.
| True faulty proportion in the lot | Approximate probability of acceptance | Who is exposed at this point |
|---|---|---|
| 1.0% | Very high, above roughly 95% | Production side exposed: a good lot occasionally rejected |
| 2.5% | High, around the level the plan is built for | Balanced, and the point the AQL nominally describes |
| 5.0% | Moderate, acceptance no longer the likely outcome | Both sides exposed, and disputes cluster here |
| 8.0% | Low | Brand exposed: a poor lot may still pass |
| 12.0% | Very low, near the limiting quality end | Brand exposed at roughly the stated consumer risk |
The shallow ends are the practical point. A plan is good at separating a 1 percent lot from a 5 percent lot and poor at separating an 8 percent lot from a 12 percent one, because by then acceptance is unlikely either way and the sample is too small to rank them.
Acceptance Number Zero and What It Really Says
Zero acceptance is often requested as though it meant zero faults, and it means something narrower: one fault found in the sample rejects the lot, which is a statement about the sample and not about the shipment.
- Zero defects in a sample of 50 bounds the true rate at roughly 5.8% at 95% confidence, not at zero.
- Achieving a bound near 1% by sampling alone needs a sample in the hundreds, which changes the cost of inspection entirely.
- Zero acceptance makes sense for critical defects, where the consequence of one occurrence is disproportionate.
- Zero acceptance on cosmetic faults converts every inspection into a negotiation about what counts as a fault.
- A better route to a low outgoing rate is a strict AQL plus rectification, rather than an unachievable acceptance number.
The second line is the arithmetic most plans never confront. Sampling is an inefficient way to prove a low rate; it is an efficient way to detect a change. Using it for the first job is what produces disappointment.
Average Outgoing Quality and Its Limit
Average outgoing quality is the proportion of faulty pieces actually shipped after rejected lots have been reworked and re-presented, and its limit is the highest that figure reaches across all possible incoming quality levels.
This is the number a brand should care about and almost never asks for. Incoming quality describes what was submitted; outgoing quality describes what arrives. Rectification is what separates them, and a plan without a rework clause has an outgoing quality equal to its incoming quality, because a rejected lot that is simply re-labelled and re-presented has not changed.
For a plan at AQL 2.5 with a sample of 50 and accept 3, the average outgoing quality limit sits in the low single digits as a percentage. That is the honest ceiling over a season, and it is a better thing to know than the AQL alone.
Capability: Why Cpk Is a Different Question
Capability asks whether a process can hold a tolerance repeatedly, and it is measured on variables rather than on attributes, which makes it a different tool used for a different purpose.
Attribute sampling
Counts pieces as faulty or sound against a defined library. This is what an AQL plan does, and it answers whether a particular lot should be released. It says nothing about whether the process that made it is stable.
Variable measurement
Measures a characteristic such as stitch pitch, panel alignment or plating thickness, and computes a capability index. It answers whether a tolerance can be held, and it needs far fewer pieces to say something useful.
Control charting
Plots a measured characteristic across the run so a drift is visible before it becomes a fault. It is the only one of the three that finds a problem while it is still inexpensive to correct.
Stitch pitch is the clearest example. Counted as an attribute, a skipped stitch is a fault. Measured as a variable, the pitch distribution tells you the machine is drifting toward the tolerance edge three hundred pieces before it produces one.
What Sampling Statistics Cannot Tell You
Four questions are regularly put to a sampling plan, and it answers none of them, which is why plans disappoint people who expected more.
- Whether the lot is free of faults: only a sample was examined, and a clean sample bounds rather than proves.
- Whether the lot will last: durability is answered by load and cycle testing, not by counting.
- Whether the materials comply: that is answered by declarations from the tannery and the finisher.
- Whether the next lot will match this one: each lot is a separate population with its own plan.
Stated plainly, a sampling plan is a rule for allocating risk on one lot. Everything else a brand wants to know comes from a different document, and the programme is only as strong as the weakest of those documents.
Putting the Statistics Into the Clause
A statistically literate clause is six lines, and each line removes one of the arguments that otherwise happens at the end of a season.
Six lines that make the plan defensible
- Standard and scheme: ISO 2859-1, single sampling, normal inspection, named rather than implied.
- Level and sample: General Inspection Level II, code letter recorded, sample size stated.
- Limits: AQL 2.5 for major, looser for minor, zero for critical, all three written separately.
- Classification: a written defect library, agreed before bulk, so the count is not negotiated later.
- Rectification: rework and re-presentation on rejection, which is what makes outgoing quality better than incoming.
- Switching: tightened after two of five rejected, normal after five accepted, reduced only on ten steady lots.
Around it, the rest of the programme is unchanged: MOQ 500 pcs per style and colourway, sampling in 6-10 working days and 12-15 for complex constructions, bulk in 35-50 days, and handover on FOB Xiamen terms. Export experience on this programme runs since 2014 across 7 production lines with about 200,000 pcs per month of capacity, delivered through our validated partner factory Quanzhou Yingming Bags.
The economics stay simple. Reworking a single closure on a piece that has reached a customer costs around USD 4.80 in parts and bench time before freight, so the cost of a plan that detects a drift early is recovered by a very small number of avoided returns.
Any plan can be quoted as a number. Asking for the operating characteristic curve, the limiting quality and the rectification rule takes one more email and tells you what the number actually delivers across a season rather than on one lot.
Get the sampling plan in writing
Tell us the quantity and the markets and we will return the sampling plan, the defect library and the switching rules before bulk begins. Request a quote.
Frequently Asked Questions
What does statistical QC decide in a bag programme?
It decides how sampling risk is allocated between the two parties when only part of a lot is examined. It does not measure how good the lot is; it states the chance a lot at a given quality level is accepted, and it does that well.
What is producer risk in a sampling plan?
The chance that a lot at the agreed quality level is nevertheless rejected, commonly around 5 percent. It is carried by the production side, and it exists in every plan because sampling without risk for either party is not possible.
What is consumer risk?
The chance that a lot at limiting quality is accepted, commonly around 10 percent. It is carried by the brand, and it is usually the larger of the two exposures because the sample is small relative to the lot.
What is an operating characteristic curve?
A plot of the probability of acceptance against the true proportion of faulty pieces in a lot. It is the whole plan drawn as a line, and reading it tells you far more than the AQL number on its own ever will.
Does zero defects in the sample mean the lot is clean?
No. With a sample of 50 and zero faults, the true rate is bounded at roughly 5.8 percent at 95 percent confidence. Proving a rate near 1 percent by sampling alone requires a sample in the hundreds, which changes the economics.
What is average outgoing quality?
The proportion of faulty pieces actually shipped once rejected lots have been reworked and re-presented. Its limit is the worst value that figure reaches across all incoming quality levels, and it is the number a brand should ask for.
Why does rectification matter so much?
Because it is the only mechanism that makes outgoing quality better than incoming quality. A rejected lot that is re-presented without rework has not changed, and a plan without a rectification clause delivers its incoming rate unchanged.
What is Cpk and how does it differ from AQL?
Cpk is a capability index computed from measured values, answering whether a process can hold a tolerance repeatedly. AQL sampling counts pieces as faulty or sound and answers whether one lot should be released. Different tools, different questions.
Why is variable measurement useful on stitch pitch?
Because a measured distribution shows drift before it produces a fault. Counted as an attribute, a skipped stitch is a fault; measured as a variable, the pitch tells you the machine is approaching the tolerance edge hundreds of pieces earlier.
What is limiting quality?
The poor quality level that a plan accepts only with the stated consumer risk, typically around 10 percent. It is the honest description of what a plan lets through, and it is the number almost nobody asks for.
How do switching rules change the statistics?
Tightened inspection after two of five rejected lots draws a larger sample at the same AQL, improving discrimination. Reduced inspection after ten steady accepted lots draws less and accepts more risk, which suits a mature style only.
What can a sampling plan not tell me?
Whether the lot is free of faults, whether it will last, whether the materials comply, and whether the next lot will match. Those are answered by full inspection, cycle testing, tannery declarations and a fresh plan per lot respectively.
People Also Ask
What is statistical QC in simple terms?
A rule for allocating risk when only part of a lot is examined. It states the chance a lot at a given quality is accepted, with producer risk near 5 percent and consumer risk near 10 percent at limiting quality.
Is AQL 2.5 the same as a 2.5% defect rate?
No. It is the quality level the plan is built around, not a permitted rate. Accepted lots can contain more or fewer faults, and the operating characteristic curve is what describes the actual behaviour.
What does a zero acceptance number really mean?
That one fault in the sample rejects the lot. It is a statement about the sample, not a guarantee about the shipment: zero faults in fifty still bounds the true rate at roughly 5.8 percent at 95 percent confidence.
Why does outgoing quality matter more than incoming?
Because outgoing quality describes what actually ships after rectification. Without a rework clause, a rejected lot re-presented unchanged delivers its incoming rate, and the plan has improved nothing at all.
When should a brand use capability instead of sampling?
When the question is whether a tolerance can be held, such as stitch pitch or panel alignment. Sampling detects a change in a lot; capability describes the process and finds drift before it becomes a fault.
Sources and further reading
Every figure on this page is either our own programme data or attributed below. Open the originals before you write them into a specification.
- ISO — International Organization for StandardizationISO Publisher of ISO 2859-1, which defines code letters, sample sizes, accept and reject numbers and the switching rules between normal, tightened and reduced inspection.
- ASQAmerican Society for Quality Reference body for acceptance sampling, operating characteristic curves and process capability indices used in quality engineering.
- NISTNational Institute of Standards and Technology Publishes the engineering statistics handbook covering acceptance sampling, OC curves and capability analysis methods.
- SATRASATRA Technology Centre Independent testing organisation for leather and accessories, covering test methods referenced in defect classification.
- Sourcing JournalSourcing Journal Reporting on inspection practice, acceptance sampling and quality disputes across the accessories trade.