Can a specification have only an upper limit?

Yes, when only a maximum is relevant.

Are specification limits estimated from process data in the same way as control limits?

No; they represent requirements, although evidence can inform their design.

Does a high Cp guarantee low nonconformance when the mean is off center?

No; centering and distribution also matter.

AI Glossary / Statistics and probability / Foundation

Specification limits

Definition

Specification limits are the numerical boundaries that define acceptable values for a specified characteristic of a product, service or process output. A lower specification limit sets a minimum acceptable value; an upper specification limit sets a maximum. Some characteristics need both, while others require only one. These limits express requirements. They are different from statistical control limits, which describe expected variation under an established process baseline.

What it means

A specification answers what the output must satisfy. It may come from design needs, customer agreements or another documented requirement. Observed process performance answers a different question: what the process currently produces. A process can be stable but routinely generate unacceptable output, or unstable while its recent values remain within specifications. Keeping these questions separate helps avoid the false conclusion that a requirement has been met simply because a process behaves consistently.

How it works

Define the characteristic, units, measurement method and relevant operating conditions. State the applicable lower or upper limit and how boundary values and measurement uncertainty are handled. Collect suitable process data and compare outputs with those requirements. If assessing capability, first examine process stability and use distributional assumptions appropriate to the data. Do not derive acceptance requirements merely by drawing lines around current observations; doing so can turn poor performance into an apparently acceptable specification without addressing the underlying need.

Three practical examples

Example 1: A manufactured component has an acceptable dimensional interval defined by its intended fit. Example 2: A service agreement states an upper response-time requirement for a specified percentile and workload. Example 3: An AI evaluation protocol defines a minimum acceptable performance metric on a named test population. These requirements concern different statistical objects. A limit on each individual item, a batch average and a percentile are not interchangeable, so the specification must say exactly what is being assessed.

Think of it this way

Think of a parking space and a car. The space defines where the car must fit; observing where a driver usually parks does not redefine the boundaries. Similarly, a process distribution describes current behavior, while specification limits describe acceptability. Improving consistency can help, but a consistently misplaced car still fails to fit within the intended space.

Worked example

Suppose a stable normal process has mean 101 and standard deviation 0.5, while acceptable individual measurements lie from 98 to 102. The specification width is four. The illustrative capability index Cp is four divided by six times 0.5, or 1.3333. The centering-sensitive index Cpk is the smaller of (102 minus 101)/1.5 and (101 minus 98)/1.5, giving 0.6667. Under the stipulated normal model, the upper limit is two standard deviations above the mean and the lower limit is six below it, so approximately 2.275 percent of output lies outside specifications. The individual three-standard-deviation baseline interval is 99.5 to 102.5. A value of 102.2 can therefore be inside that illustrative control interval while exceeding the upper specification. Capability indices and probability statements require the stated stability, distribution and measurement assumptions.

Illustrative case study

In this fictional case study, Silverline Components reports a stable production chart and initially assumes quality is satisfactory. A separate review shows that the process is centered too close to the upper dimensional requirement. The team preserves the customer specification, investigates how to improve centering and verifies the measurement procedure. Managers receive separate reports on process stability and conformance, preventing a stable pattern from being mistaken for an acceptable one. Follow-up data assess whether the improvement lasts.

Common misconceptions and mistakes

Common mistakes include replacing specifications with control limits, treating Cp as a complete assessment of an off-center process, and assuming that a few conforming observations prove future capability. Another mistake is applying normal-based capability calculations without checking whether their assumptions are reasonable. Specifications also need clear scope: an average response time below a threshold does not establish that nearly all individual responses satisfy it. Rounding and measurement uncertainty can matter near a boundary and should follow the stated decision rule.

Business use

Businesses use specifications to communicate expectations, assess delivery and identify where process improvement is needed. A useful specification is measurable and tied to an actual requirement. Record who owns it, which population or operating conditions it covers and how changes are approved. For AI systems, a performance threshold should be linked to a defined evaluation procedure and relevant groups rather than a context-free headline score. Meeting a numerical limit supports only the requirement actually measured, not every dimension of product quality.

Check your understanding

What is the specification interval in the example?

Answer: 98 to 102.

Why is Cpk smaller than Cp?

Answer: The mean is closer to the upper requirement, so the available margin is unequal.

Can 102.2 be within the displayed control interval yet fail the specification?

Answer: Yes.

Would redefining the upper limit to match the process automatically solve the customer problem?

Answer: No; the underlying requirement still needs to be satisfied.

Related terms

Control limits
Related terms include tolerance
conformance
process capability
Cp
Cpk
measurement uncertainty and acceptance criteria
while a limit marks an acceptability boundary

Learn next

Learn next how capability indices combine process spread and centering under explicit assumptions. Compare requirements on individual values, averages and percentiles. Then examine how measurement uncertainty and sampling affect a conformance decision, and keep the evidence for stability separate from the evidence for meeting specifications.

Sources

Sources: NIST/SEMATECH Engineering Statistics Handbook, Glossary, https://www.itl.nist.gov/div898/handbook/glossary.htm; What is Process Capability?, https://itl.nist.gov/div898/handbook/pmc/section1/pmc16.htm; NIST Dataplot, CPL, https://itl.nist.gov/div898/software/dataplot/refman2/auxillar/cpl.htm. These primary institutional sources define requirements and capability measures. The numerical process example is original and independently checked; the case is fictional.

Previous: Specification error Next: Specificational coding