Ask a room of engineers how many parts to pull for a validation test and you will hear "thirty" with remarkable confidence. I have sat through that exchange in supplier reviews on two continents, and almost nobody can tell me where the figure comes from. The honest answer is that thirty is a rule of thumb from classical statistics, tied to the point where the t-distribution begins to resemble the normal distribution closely enough that some textbooks stopped drawing the distinction. It was never a quality planning tool.
Thirty has nothing to say about defect rates, reliability targets, or the risk you accept when the sample passes. It persists because it feels substantial without being painful: it fits in a tray and sounds like diligence in a report. But a sample size must be derived from what you need to claim, not from what sounds convincing in a meeting. When a customer asks for a reliability statement at a given confidence level, thirty is almost always either far too few or wastefully too many, and which one it is depends entirely on the mathematics you never did.
Three frameworks actually govern sample size in manufacturing practice: power analysis for variable data, confidence-reliability logic for attribute data, and the c=0 acceptance plans that tie the two together into something you can defend in an audit. None of it is exotic. All of it is routinely skipped in favour of the magic number.
Power Analysis: What You Never Did With Thirty
Power analysis answers a deceptively simple question: if the defect or shift you fear actually exists, what is the probability your sample will detect it? That probability is power, and it depends on the size of the effect, the variation in the process, the significance level you set, and the sample size. Reduce any input and you must compensate in the others. Thirty parts only gives you adequate power against effects that are large relative to your process variation, and in a capable modern process a meaningful shift is often small precisely because the noise is small.
Consider a dimension with tight bilateral tolerances. To detect a mean shift of a fraction of the tolerance width with reasonable power, you may need a sample several times larger than thirty. Test thirty parts, see nothing, and declare the process safe: you have proven nothing except that a large disturbance is absent. This failure mode is common in PPAP-style capability studies, where engineers run the minimum sample, get an unremarkable result, and mistake absence of evidence for evidence of absence. The customer's statistician, if they have one, sees through it immediately.
Working backwards is the discipline I insist on. Before testing, write down the smallest effect that matters commercially, whether a mean shift, a proportion of nonconforming parts, or a change in torque retention after cycling. Then fix the power you demand, typically set high for validation work, and compute the required sample size. The standard is the Neyman-Pearson framework, implemented in any competent statistical package. The point is not the arithmetic; the point is that the effect size is an engineering decision, made in writing, before the data exists.
Deriving a sample size instead of guessing one
- 01Define the effectWrite down the smallest shift or defect rate that matters commercially.
- 02Set power and alphaFix detection probability and significance level appropriate to validation.
- 03Compute nSolve for sample size using the Neyman-Pearson framework.
- 04Test and interpretA clean result now carries a checkable claim, not a vague assurance.
Confidence and Reliability: The Language of Attribute Testing

When the outcome is pass or fail, whether a leak test result, a functional check, or a visual defect found or not found, you are in attribute territory, and the governing logic is the confidence-reliability statement. You want to say: I am 95 percent confident that at least 99 percent of the population conforms. Notice what that sentence demands. It says nothing about averages. It makes a claim about the worst-case fraction of the lot, and the sample size needed to support it with zero observed failures is governed by the binomial distribution, not by any rule of thumb.
Here is the uncomfortable part for devotees of thirty. If you test thirty parts and see no failures, the binomial mathematics tells you that with 95 percent confidence you can only claim the defect rate is no worse than roughly one in ten. One in ten would embarrass anyone in automotive. To support a 95/99 statement with zero failures, the required sample runs into the hundreds. The exact figure comes from the reliability formula that sets the confidence bound at zero failures, and every quality engineer should be able to derive it, because it is the single most underappreciated number in our field.
The practical consequences are real. Where reliability requirements are severe, as with aerospace fasteners, medical-adjacent components, or safety-critical welds, attribute sampling alone becomes economically impossible, and that pushes you towards variables measurement instead. Measuring a continuous characteristic and knowing its distribution lets a much smaller sample support a much stronger claim, because every reading carries information, not just a binary verdict. When a supplier tells me they validated a safety characteristic with a few dozen pass/fail checks, I know the claim and the evidence do not match.
c=0 Plans and the Logic of Zero Acceptance
C=0 sampling plans, zero-acceptance plans in the tradition associated with Nicholas Squeglia, were built on an observation that still offends classical statisticians: accepting a lot after finding a defect or two in the sample feels wrong to customers, and the discrimination of small-sample plans is poor anyway. A c=0 plan sets the acceptance number at zero, so any defect in the sample rejects the lot. In exchange, the sample size needed for a given lot tolerance is smaller than a corresponding plan that permits defects, and the plan's behaviour at the operating-characteristic level is sharply asymmetric: gentle on excellent lots, brutal on marginal ones.
The mathematics behind c=0 plans connects directly to confidence-reliability logic. With zero acceptance, the sample size and the confidence level are two sides of the same binomial coin: choosing a sample size is choosing what population defect rate you can rule out when the sample comes back clean. I treat the Squeglia tables not as a mysterious lookup but as pre-computed binomial confidence statements, re-indexed by lot size. Once you see that, you stop asking what the table says and start asking what defect rate you are implicitly accepting, which is the question an auditor should be asking you.
Choosing a sample size is choosing what population defect rate you can rule out when the sample comes back clean.
C=0 plans have genuine weaknesses, and honesty about them builds credibility. A single random defect, whether a dropped part, a mislabelled bin, or a marginal gauge, condemns an otherwise good lot, so the plans punish processes with occasional, innocuous slip-ups. They also invite abuse when combined with rework loops: reject the lot, screen it, resubmit, repeat, until a clean sample appears by luck. That is not sampling; that is dice-rolling with extra steps. The plan must sit inside a disciplined response to any detected failure, escalating to containment and root cause rather than a quiet re-inspection.
The Inputs That Actually Drive the Number
Sample size is an output. The inputs are where engineering judgement lives, and there are four that matter most. First, the population definition: are you sampling a shift, a lot, a tool set, a cavity? A multi-cavity mould treated as one population hides cavity-specific defects in the averaging; a stratified sample per cavity is often the correct design, and it multiplies your count. Second, the characteristic's distribution. Variables methods that assume normality fall apart on skewed or bounded data such as flatness, runout, or times-to-failure, and the sample must either grow or the method must change to something distribution-free.
Third, measurement discrimination. If your gauge cannot resolve a tenth of the variation you care about, no sample size rescues you; you will measure the gauge, not the part. This is why MSA belongs upstream of any sample size discussion, not after it. Fourth, risk allocation. Who bears the consequence of a missed defect, and how badly? A cosmetic characteristic on an interior trim panel and a brake component's shear strength do not merit the same confidence-reliability pair, and pretending they do is how organisations burn money on one and court disaster on the other.
One more input people forget: the expected defect rate itself. If you anticipate a low but non-zero rate, a zero-failure plan needs to be sized against that expectation, and you should plan what happens when the first failure arrives. A test plan without a pre-decided reaction to failure is a plan that will be improvised under pressure, and improvisation under pressure is where bad statistics and bad quality decisions are born.
What clean samples actually certify
Common Failure Modes I Keep Seeing
The most frequent error is the transplanted rule: someone carries a sample size from a previous programme into a new one without checking whether the requirement, the variation, or the measurement method survived the journey. Close behind is the sample of convenience, pulling thirty consecutive parts at the end of the run, from the cavity nearest the operator, after the process had settled. A sample that does not represent the population over its full variation supports no statement about that population, whatever the arithmetic says. Randomisation and coverage are not optional garnish.
Then there is silent re-use of data. A capability study from the qualification run gets cited two years later as evidence for a process that has since changed tooling, material lots, and cycle time. Sample size determination includes a validity window, and the assumptions behind the number must be re-checked whenever the process changes. The number is attached to a process state, not to a part number in perpetuity.
Finally, watch for the confidence-reliability mismatch in supplier documents: a requirement of 95/99 quoted above a test report showing a few dozen attribute checks and no failures. That gap between claim and evidence is not a technicality. In a safety context, it is the finding that ends the review, and it should be caught by your own document review long before an auditor or customer finds it for you.
Making the Number Defensible
My standing advice to quality engineers is simple: never present a sample size without the sentence that justifies it. "Thirty parts, zero failures, therefore ninety-five percent confident the defect rate is below one in ten" is defensible even when the conclusion is unwelcome, because the reasoning is transparent and the mathematics is checkable. A bare "we tested thirty" is a superstition wearing a lab coat. The derivation, the assumed effect size or reliability target, and the reaction plan belong in the test report itself, not in someone's memory.
Customers respond well to this candour, in my experience, because it invites a real conversation. If the implied risk is too high, the requirement or the method changes: shift to variables measurement, tighten process control upstream, or negotiate a different confidence-reliability pair appropriate to the characteristic's severity. That conversation is quality engineering. The magic number forecloses it.
Retire thirty as an answer and keep it only as a starting point for the question it should always provoke: thirty relative to what? Every sample size in your test plans encodes a risk decision, whether or not anyone made it deliberately. The choice is between making that decision explicitly, in writing, before the data exists, or discovering it during an audit after the data has already lied to you.
