PSLE Science Reality Lab Vol No.252 | “Cpk = 1.33” — Does One Number Guarantee Every Product Meets the Limits?

Stable internal ID: PSLE-SCI-REALITY-0252

Wait, what? A factory quality report has a large green badge: Cpk = 1.33. Someone points at it and says, “Great. That proves every product coming off the line is inside specification.”

That is a much stronger claim than the number itself earns.

NIST describes process capability as a comparison between the natural variation of a stable process and its specification limits. Cpk is one capability index. It uses information about the process centre, its spread and the nearest specification limit. It is not a percentage score, and it is not a certificate that every individual item is good.

This looks advanced, but the learner job is very PSLE Science: identify what was measured, what the number represents, what assumptions sit underneath it, and how far the conclusion can travel.

Quick Answer

No. Cpk is a process-capability index. Under suitable conditions, it summarises how the location and variation of a process compare with engineering specification limits. It does not mean 1.33%, 133% quality, 99% accuracy or “zero defects guaranteed.”

Before using Cpk, ask whether the process is stable, whether the data are suitable for the calculation, whether the assumed distribution is reasonable, whether the sample is large enough, whether the measurement system is trustworthy and whether the specification limits are the correct limits for the product.

The Owned Learner Job

This Reality Lab owns one narrow communication job: how to evaluate a manufacturing report, infographic or supplier claim that presents Cpk as if one capability index were a direct guarantee about every product.

It does not own statistics, standard deviation, normal distributions or manufacturing engineering as standalone subjects. It also does not set a universal “good Cpk” threshold. Different industries and applications can have different requirements. The evidence habit is to read the index according to its definition and assumptions.

Build a Child-Sized Factory Example

Imagine a factory makes small rods intended to be between 49.5 mm and 50.5 mm long. Those are the lower and upper specification limits.

The machine does not produce every rod at exactly 50.000 mm. Real processes vary. One rod may be 49.97 mm, another 50.08 mm, another 49.91 mm. The factory measures a sample and studies the centre and spread of those measurements.

Cpk asks, in simplified language:

How comfortably does the process distribution fit inside the specification window, especially on the side closest to a limit?

That is a very different question from “Did we inspect every rod?”

Three Objects: Specification Window, Process Cloud, Capability Index

1. The specification window

This is the acceptable range defined for the product: for example, 49.5 mm to 50.5 mm.

2. The process cloud

This is the distribution of actual production measurements. Is it narrow or wide? Is it centred or shifted toward one limit? Is it stable over time?

3. The capability index

Cpk compresses information about the process centre, spread and nearest specification boundary into one index. Compression is useful, but one number can hide details. The scientist must know what went into the number.

Observed, Calculated and Claimed

Observed: the measured sizes, masses, strengths or other product values in a sample.
Calculated: statistics describing the centre and spread, then a capability index such as Cpk.
Claimed: statements such as “the process is capable,” “all products pass,” or “quality is guaranteed.”

The calculated index can be valid while the final marketing sentence still overreaches.

Why Stability Comes First

NIST emphasises that process capability is assessed for a stable process. Why?

Suppose a machine slowly drifts longer every hour because a tool is wearing. If we mix early and late measurements into one calculation, a single capability index may hide the time pattern. Tomorrow’s process may not behave like yesterday’s process.

Stability asks whether the process behaves with reasonably constant statistical properties over time. Capability asks how that process compares with the specification limits. Those are related but different jobs.

Worked Case 1: Narrow but Off-Centre

Machine A produces rods with very little variation, but its average is close to the upper specification limit. The process cloud is narrow, yet it is crowded against one boundary.

A learner who looks only at “small variation” may think the process is excellent. Cpk cares about the nearest limit, so poor centring can reduce capability even when variation is small.

The evidence lesson is broader than Cpk: precision around the wrong centre is not the same as meeting the intended target comfortably.

Worked Case 2: Centred but Too Wide

Machine B is centred beautifully at 50.0 mm, but its measurements spread widely. Some values approach or cross both specification limits.

A centred average does not prove individual products are consistent. The spread matters.

Worked Case 3: The Tiny Sample

A factory measures only five parts on a quiet morning and calculates a very high Cpk. It then prints “world-class capability” on a brochure.

The problem is not that the formula suddenly becomes evil. The problem is evidence quantity and representativeness. NIST notes that capability estimates need enough independent data; its handbook gives about 50 observations as a common practical idea for normal-based capability estimates, while actual requirements depend on the analysis.

Five unusually similar parts may not reveal the normal variation of a long production run.

Worked Case 4: Two Machines Hidden Inside One Number

Machine C makes parts centred slightly low. Machine D makes parts centred slightly high. A report mixes both machines into one dataset and calculates one Cpk.

The combined distribution may look wider or oddly shaped because it contains two different processes. A single index can hide the fact that the data came from different sources.

Before trusting a summary statistic, ask what population was combined.

Worked Case 5: Measurement Error Inflates the Spread

Suppose the real rods are fairly consistent, but the measuring tool is noisy. Repeated measurements of the same rod vary noticeably. The observed process spread now contains both production variation and measurement variation.

A capability calculation cannot magically separate those sources. Measurement quality must be good enough for the job.

Cpk Is Not 133%

The number 1.33 looks like something that should become 133%. Resist that reflex. Cpk is a dimensionless ratio-like index built from distances to specification limits and process variation. It is not a percentage scale.

Likewise:

Cpk = 1.33 does not mean 1.33% defective.
It does not mean 98.67% good.
It does not mean 133% accurate.
It does not guarantee zero defects.

Under particular statistical assumptions, capability indices can be related to expected out-of-specification fractions. But that calculation depends on the model and assumptions. The index itself is not a defect percentage printed in disguise.

The Distribution Assumption Matters

NIST’s common Cpk formulas assume normally distributed process values. If the real distribution is strongly skewed, has two peaks, contains time trends or includes special causes, a normal-based capability calculation can misrepresent the process.

A graph of the data should accompany the number whenever possible. Never let a single index replace looking at the actual distribution.

Specification Limits Are Not Control Limits

Specification limits come from product requirements: what values are acceptable for the item. Control limits are statistical boundaries used to watch process behaviour over time.

They are not interchangeable. A stable process can still be badly positioned relative to product specifications, and a capable-looking process may become unstable later.

What Evidence Strengthens a Cpk Claim?

The process has first been shown to be stable over a relevant period.
The sample contains enough independent observations.
The data distribution is checked rather than assumed blindly.
The measurement system is suitable for the required tolerances.
The correct upper and lower specification limits are used.
Data from different machines, products or conditions are not mixed carelessly.
The report shows the histogram, time plot or other context behind the index.
The conclusion is about process capability, not a universal guarantee about every unit.

What Weakens It?

A Cpk value with no sample size or time period.
A drifting or unstable process.
A tiny convenience sample.
A strongly non-normal or mixed distribution treated as normal without checking.
Unreliable measurement tools.
Unknown specification limits.
A claim that converts Cpk directly into “percent good” without explaining assumptions.
A guarantee that every future product will pass.

How Far Can the Conclusion Travel?

A careful statement is:

For this stable process and dataset, the calculated Cpk summarises how the process centre and variation compare with the stated specification limits under the model assumptions used.

That is different from “every part passed,” “the process will stay this way forever,” or “the quality is 133%.”

Tempting but Invalid Reasoning

“Cpk is above 1, so every item must pass.” A capability index is not an inspection certificate for every item.
“1.33 means 133% quality.” Cpk is not a percentage scale.
“A high Cpk means the process is stable.” Stability should be established separately.
“The average is on target, so capability must be good.” Spread matters too.
“The sample Cpk is high, so all future production is guaranteed.” Future performance depends on whether the process remains in the same state.

PSLE-Style Transfer Case: Bottle Fill Volume

A bottling machine is intended to fill between 495 mL and 505 mL. A quality report uses measurements from 60 bottles and calculates a capability index. The next day, a loose machine part changes the filling behaviour, but the old Cpk badge remains on a dashboard.

Can the old number guarantee the new day’s bottles?

No. The capability index described the process represented by the earlier data. If the process changes, new evidence is needed to show that the earlier capability still applies.

Explained Practice

Practice 1

A report says Cpk = 1.50. Can you call that 150% accurate?

Answer: No. Cpk is a process-capability index, not an accuracy percentage.

Practice 2

A process has a narrow spread but its mean is close to the upper specification limit. What should you notice?

Answer: Low variation alone is not enough. The process is off-centre and has less room before crossing the nearest specification limit.

Practice 3

Why look at a time plot or control chart before relying on a capability index?

Answer: Capability calculations assume the process behaviour being summarised is stable enough to represent a consistent process.

Practice 4

Why is a large sample from one machine not automatically evidence for a different machine?

Answer: The two machines may have different centres, variation or causes of error. Evidence belongs first to the process that produced it.

Delayed Independent Return: Draw the Window and the Cloud

Tomorrow, draw two vertical lines for the specification limits. Then sketch three process clouds: narrow and centred, narrow but shifted, and wide but centred. Without using formulas, explain which features make each case more or less capable. If you can do that, the index has become a scientific idea rather than a mysterious badge.

Parent and Tutor Teaching Guide

Use a simple target game. Draw an acceptable band on paper and place dots representing manufactured parts. First keep the dots tightly centred. Then shift the whole group toward one edge. Then widen the group. Ask what changed: centre, variation or both.

Only after the learner can see those three ideas should you introduce the label Cpk. This prevents formula memorisation from replacing understanding.

The transferable habit is: when one quality number appears, reconstruct the data distribution and limits hiding behind it.

Route to Existing Canonical PSLE Science Owners

Authoritative Sources

Quiet Return

Cpk is useful because it compresses a lot of process information. That is also why it can be overread.

When one capability number appears, unfold it: Which process? Which limits? How stable? How much variation? How centred? Which sample? Which assumptions? Only then decide what the number has earned.

Small Group Tutorials