Process capability: proving a process can hold a tolerance without being watched

Stability first, capability second

An index calculated from a process that was shifting during the study describes an average of several different processes and predicts nothing. The order of work is therefore fixed: establish that the process behaves consistently over time, then measure how its output sits against the tolerance. Skipping the first step produces the familiar situation where a supplier submits strong capability figures at approval and then misses the same characteristic in production. Nothing was falsified. The study simply measured a period during which the process happened to be well behaved, and the underlying instability was never examined.

Centring and spread fail differently and need different fixes

A process can be tight but sitting off nominal, or centred but too variable, and the two demand opposite responses. An off-centre process is usually a setting, an offset, a fixture or a tool datum, and it can often be corrected in an afternoon at almost no cost. Excess spread is inherent to the equipment, the material or the method, and reducing it means changing something substantial. Reading capability without separating the two leads plants to launch improvement projects against processes that merely needed adjusting, which wastes effort and discredits the exercise.

What the number quietly assumes

The usual indices assume the output follows a symmetric bell-shaped distribution, and many real characteristics do not. Anything bounded at zero, such as flatness, runout or concentricity, is skewed by its nature. Anything sorted before measurement is truncated. Anything produced by several cavities or spindles is a mixture of distributions rather than one. Applying the standard formula to those cases produces figures that look reassuring and mean nothing. Plotting the data and looking at its shape before calculating anything is a habit worth more than any amount of arithmetic sophistication.

Capability evidence should change what you check

The commercial return on this work comes from acting on it. A characteristic demonstrated to sit comfortably inside its tolerance over a sustained period does not need the checking rhythm applied to one that drifts, and reducing that effort is how the study pays for itself. Equally, a characteristic that cannot be held requires either full verification, a process change, or a conversation with the customer about the tolerance. What should not happen is capability being calculated for a submission, filed, and inspection continuing exactly as before regardless of what the data showed.

Sample period matters more than sample size

Data gathered from a single run on one shift with one material lot describes a narrower world than production will occupy. Longer-term evidence collected across shifts, operators, tool changes and material batches always looks worse, and it is the honest figure, which is why customers frequently ask for both. When a supplier's capability appears strong and their delivered quality does not, the gap is almost always in what the study spanned rather than in the arithmetic. Ask what period the data covers before questioning anything else. State the span on the report itself, since a figure quoted without its period is an assertion the receiving engineer has no way to evaluate.

Frequently asked questions

What is the difference between short-term and long-term capability figures?

Short-term studies cover a limited run under consistent conditions and show what the process can do at its best. Long-term figures span shifts, operators, tool changes and material lots, and show what it actually does. The second is nearly always the weaker number and the more useful one. A large gap between them points at sources of variation that come and go, which is a more productive line of investigation than trying to improve the machine itself.

Our capability figures look good but customers still find defects. Why?

Several explanations recur. The study may have covered a period that was not representative, the measurement system may be adding or masking variation, the characteristic causing complaints may not be the one being studied, or the defects may arise from sporadic events such as mix-ups and handling damage that a distribution never describes. Capability speaks about ordinary variation in one measured feature; it has nothing to say about the incidents that produce most complaints.

Can capability be assessed for a characteristic that is judged rather than measured?

Not with the usual indices, which need measured values. For pass-or-fail characteristics the equivalent evidence is the observed defect rate over a defined period, together with confirmation that the judgement itself is consistent between inspectors. Where a customer asks for capability on a visual characteristic, the productive response is to offer a defect rate and an agreement study rather than to force the data into a formula that does not fit it.