What does a clean sample of 14 actually prove?

That the defect rate is probably below about 19%, at 95% confidence. That sounds disappointing and it is exactly right: with zero defects found, the upper bound is 1 − (1 − C)^(1/n), and small n gives a loose bound. It is enough to detect a lamp out or a wrong-speed belt, which is a real and worthwhile job. It is not enough to support a claim about parts per million.

How many samples would I need for my PPM target?

Roughly ln(1 − C) ÷ ln(1 − p), for 200 PPM at 95% confidence, about 15,000 parts inspected with no defect found. On a 2,400-part shift that is more than six shifts of complete inspection. This is why PPM-level quality on a UV line is demonstrated by process control rather than by sampling: no achievable sampling plan reaches it.

Does higher confidence mean a better result?

It means a stronger claim, and from the same evidence a stronger claim must be a weaker bound. At 99% confidence the same 14 samples only rule out 28% rather than 19%. Nothing about the process changed. You are simply asking for more certainty from the same fourteen parts, and the arithmetic charges you for it.

If sampling cannot reach PPM, what should?

Monitoring the cause. On a UV line the cure depends on dose, and dose can be measured continuously: radiometer readings on a schedule, dose on the part with a puck or strips, lamp hours, belt speed and lamp-out alarms. Those cover every part instead of one in two hundred and they detect a drifting lamp days before a small sample could. Sampling then confirms the link between the monitored quantity and a cured part, which is a job it can do.

What if a sample fails?

The bound on this page no longer applies. It assumes zero defects, and one failure moves the estimate a long way. More importantly, a cure defect found in a sample of 14 implies a rate high enough to be a process event rather than a quality statistic: investigate the lamp, the speed, the material and the fixture rather than pulling more samples.

Why keep the inspection efficiency divisor?

Because it is a fair allowance for a real effect. Rub tests and gel-content preparations can be spoiled or read inconclusively, so pulling more than the nominal count is how you end up with the intended number of valid results. It inflates the pull without strengthening the statistics. The bound uses the pull, and if efficiency is poor the effective sample is smaller than it looks.

UV Curing calculator

UV Cure Validation Sample Size, and What It Actually Proves

Size a cure validation sample, and find out what it can actually support. Enter the shift's production volume, the plant's sampling rate, the inspection efficiency, and the confidence level. The calculator returns the sample pull, the earlier version's answer, unchanged, plus the defect rate that sample rules out if every part passes, and the number of samples a 1% demonstration would need. Those two rows exist because a sample size means nothing without them: 14 good parts sound reassuring, and they are consistent with a process running at nearly one in five bad.

What this calculator does

Size a cure validation sample from a sampling rate, and state the defect rate that sample can actually rule out if every part passes.
Use it for sizing a routine cure validation pull, stating what a clean validation actually demonstrates, showing why a PPM target cannot be verified by sampling, comparing sampling plans at different confidence levels, justifying dose monitoring as the primary cure control.
Size a cure validation sample from a sampling rate, and state the defect rate that sample can actually rule out if every part passes.

Formula used

Raw samples = production volume × sampling rate ÷ 1,000
Recommended sample pull = CEILING(raw samples ÷ inspection efficiency)
Defect rate ruled out (zero defects found) = 1 − (1 − confidence)^(1 ÷ n)
Samples to demonstrate 1% = CEILING(log(1 − confidence) ÷ log(0.99))
Share of production inspected = sample pull ÷ production volume × 100

Inputs explained

Production volume per shift: Parts cured in the shift the validation covers.
Sampling rate: The plant's sampling convention. It is a policy, not a statistical derivation: the rows below say what it buys.
Inspection efficiency: Share of pulled samples that yield a usable result, after spoiled preparations and inconclusive reads. Pull more so that enough are valid.
Confidence level: Confidence for the upper bound on defect rate. 95% is the usual convention; a higher level makes the claim stronger and the bound weaker.

How to use the result

Best suited to sizing a routine cure validation pull, stating what a clean validation actually demonstrates, showing why a PPM target cannot be verified by sampling, comparing sampling plans at different confidence levels, justifying dose monitoring as the primary cure control.
The ruled-out rate assumes zero defects in the sample. Find one and the bound moves a long way in the wrong direction. This page does not compute that case. Assumes a perfect inspection. A test that misses undercure some of the time inflates the apparent confidence, and cure defects are exactly the kind that hide from a quick check. Says nothing about WHEN the defects occur. A sample pulled at the start of a shift proves little about a lamp that drifts by the end of it. Treats the population as homogeneous, when a shift often contains several materials, lamp states and operators. Attribute sampling only: a pass/fail result carries far less information than a measured dose would from the same part.

Common questions

What does a clean sample of 14 actually prove? That the defect rate is probably below about 19%, at 95% confidence. That sounds disappointing and it is exactly right: with zero defects found, the upper bound is 1 − (1 − C)^(1/n), and small n gives a loose bound. It is enough to detect a lamp out or a wrong-speed belt, which is a real and worthwhile job. It is not enough to support a claim about parts per million.
How many samples would I need for my PPM target? Roughly ln(1 − C) ÷ ln(1 − p), for 200 PPM at 95% confidence, about 15,000 parts inspected with no defect found. On a 2,400-part shift that is more than six shifts of complete inspection. This is why PPM-level quality on a UV line is demonstrated by process control rather than by sampling: no achievable sampling plan reaches it.
Does higher confidence mean a better result? It means a stronger claim, and from the same evidence a stronger claim must be a weaker bound. At 99% confidence the same 14 samples only rule out 28% rather than 19%. Nothing about the process changed. You are simply asking for more certainty from the same fourteen parts, and the arithmetic charges you for it.
If sampling cannot reach PPM, what should? Monitoring the cause. On a UV line the cure depends on dose, and dose can be measured continuously: radiometer readings on a schedule, dose on the part with a puck or strips, lamp hours, belt speed and lamp-out alarms. Those cover every part instead of one in two hundred and they detect a drifting lamp days before a small sample could. Sampling then confirms the link between the monitored quantity and a cured part, which is a job it can do.
What if a sample fails? The bound on this page no longer applies. It assumes zero defects, and one failure moves the estimate a long way. More importantly, a cure defect found in a sample of 14 implies a rate high enough to be a process event rather than a quality statistic: investigate the lamp, the speed, the material and the fixture rather than pulling more samples.

Last reviewed 2026-08-25.