Process capability: qualify the data before Cp and Cpk

A strong weekly yield can hide a feature drifting toward its upper limit. Keep specification revisions, sampling scope, and stability evidence beside the index.

Kartik Deshpande··7 min read

If the weekly yield report looks healthy while inspection records show oversized shafts, start the process capability review with that diameter and the process that produced it. Fix the specification revision, select representative measurements, and establish stability before interpreting Cp or Cpk. A plant-wide acceptance rate can combine products and features that face different limits, so it cannot answer whether this diameter has enough room before its upper specification.

ASQ describes capability estimates as comparisons between process spread and specification limits for continuous measurements. Use them to qualify a defined production process under stated conditions. Keep the observed acceptance count alongside the study; first pass yield answers the separate question of how much output passed without rework.

Fix the characteristic and specification revision

Start with a measurable characteristic, such as shaft diameter at a defined inspection location. Record the part revision, unit, measurement method, target, and approved lower and upper specification limits. Assign the engineering or customer document that authorizes those limits. When a drawing changes during the study period, retain the effective revision on each measurement instead of judging all history against the latest drawing.

Keep specification limits separate from control limits. NIST’s control chart description derives the latter from the behavior expected of an in-control process. Specifications state what the product must meet. Copying chart limits into the acceptance fields would let observed process variation redefine the requirement.

For a characteristic with only an upper requirement, retain a one-sided specification. NIST defines an upper capability index, Cpu, using the distance between the mean and that limit. Do not invent a lower limit to make a two-sided Cp calculation possible, and do not substitute a target for a missing limit.

Build a study population that preserves process history

Join inspection results to production order, operation, machine or cavity, material lot, and measurement time where those fields exist. Preserve the inspection-result ID and the original value, including failures. A spreadsheet containing only released material removes the observations needed to understand the upper-limit problem. If the source retains only pass/fail results, retrieve the measured values before applying these continuous-data indices.

Define the production conditions the conclusion will cover. A study of one machine and one setup cannot establish capability for every machine making the part. Conversely, pooling cavities with different centers may conceal which cavity needs attention. Inspect those groups before deciding whether a shared process estimate is defensible. Retain the grouping decision so a reviewer can reproduce the same population.

ASQ’s sampling discussion shows why repeated samples across a run can represent production better than one consecutive group. Design the study to cover the conditions in its scope, including relevant setup or material changes. Agree on rational subgroups (observations made under comparable short-term conditions) and the variation estimator with the quality owner before calculating indices.

Verify that the measurement method can resolve the differences the study needs to evaluate. Retain calibration status and the applicable measurement-system assessment with the dataset; unresolved bias or repeatability concerns need review before a machine adjustment follows from the numbers. The shop floor data collection workflow covers the wider task of preserving production context at capture.

Establish stability before interpreting the index

NIST’s stability assessment requires a constant process distribution over time, including its mean and variance, before discussing capability. Keep measurements in production order and review an appropriate control chart with the quality engineer. A histogram removes time order, so a single bell-shaped display does not establish that the process stayed stable.

Investigate signals against the production history. Mark a tool replacement, setup change, or measurement-method change where it occurred and retain the original records. If a documented cause creates a distinct operating period, define and analyze that period on its own terms. Deleting inconvenient observations until the chart looks acceptable produces a different study population without explaining the original failures.

Check distributional fit after settling the process scope. ASQ states that Cp and Cpk are sensitive to normality assumptions and describes transformations or alternative distributions for stable, non-normal data. Ask the quality engineer to choose and document the method. A normal-model calculation on skewed data does not establish the predicted tail risk.

Keep qualification beside the result

Index alone

Mixed runs, current limits, hidden exclusions

Qualified study

Defined scope, stable process, retained method

A reported index needs the population and method that produced it so a later review can test the same conclusion.

Calculate Cp and Cpk against the same assumptions

NIST gives the sample formulas for Cp and Cpk as Cp = (USL - LSL) / (6 × s) and Cpk = min[(USL - mean) / (3 × s), (mean - LSL) / (3 × s)]. USL and LSL are specification limits; s is the estimated process standard deviation under the declared method. Cp compares width, while Cpk also accounts for the mean’s distance from the nearer limit.

For a worked arithmetic example, assume a shaft diameter with LSL = 9.90 mm, USL = 10.10 mm, mean = 10.06 mm, and s = 0.02 mm. These are invented teaching inputs, not measured performance. Assume representative continuous measurements, an adequate measurement system, a stable approximately normal process, and an agreed variation estimator. The inputs alone provide no evidence that those conditions hold.

CalculationSubstitutionResult
Cp(10.10 - 9.90) / (6 × 0.02)1.67, rounded
Upper-side component(10.10 - 10.06) / (3 × 0.02)0.67, rounded
Lower-side component(10.06 - 9.90) / (3 × 0.02)2.67, rounded
CpkSmaller of the upper-side and lower-side components0.67, rounded

The specification width is 0.20 mm, and the estimated six-standard-deviation spread is 0.12 mm. Yet the mean sits only 0.04 mm below the upper requirement. The upper side therefore controls Cpk even though Cp exceeds one. This calculation identifies an off-center condition under the assumptions; it does not estimate a measured reject rate or authorize release of a lot.

Label the variation method beside the output. A within-subgroup estimate and a standard deviation across the whole reporting window can capture different sources of variation. Replacing one with the other while keeping the report label unchanged prevents comparison with the prior study. Retain the software settings or calculation definition, sample count, time span, and subgroup structure with the result.

Bring the inspection evidence into one review

Start with permitted inspection extracts, production references, and the specification revisions used at measurement time.

Turn the finding into a documented quality decision

Review the failed measurements as well as the index. In the opening scenario, inspect oversized results by machine, tool age, and measurement time before deciding whether to change a setup or investigate the gauge. NIST describes centering and reducing variation as separate improvement needs for stable processes that fail requirements. The corrective action should address the supported cause.

Use the customer’s qualification requirement and the quality plan to set the acceptance rule before seeing the result. Record the required index, estimation method, evidence of stability, and treatment of sampling uncertainty. NIST discusses confidence intervals for capability indices because an estimate varies with the sampled data. No chosen Cpk threshold guarantees that every future unit will conform.

After an approved change, collect evidence under the revised conditions and preserve the earlier study. Keep containment, lot disposition, and corrective-action approval with their owners. If that handoff lacks controlled records, the quality management system software comparison covers inspection and corrective-action workflows to evaluate.

Where Permute fits in a capability review

Permute sells connected business data, custom applications, and governed analysis. For this review, we can combine permitted inspection extracts from Excel or Google Drive with production references from SAP or NetSuite. We retain source references so an unmatched part revision remains a review item instead of receiving a guessed specification.

We can apply your declared characteristic definitions, units, revision mappings, and study filters across repeated analyses. The quality owner supplies the sampling and statistical method. Keeping those rules explicit lets the next review use the same definition or show which approved rule changed.

A quality lead can ask which measurements entered the study, which were excluded, and which specification revision each used. We can return the supporting source records their permissions allow. That governed evidence can also support questions in Claude, ChatGPT, or Copilot while the engineering decision remains with the quality team.

Keep statistical approval and process changes with their owners

This analysis reads approved source data. We do not establish gauge suitability, certify process stability, approve a statistical model, or release product on the basis of a chat answer. Missing timestamps, specification history, or failed-result records remain limits on the conclusion. A connected dataset cannot recover measurements the source never captured.

This workflow does not promise machine adjustments, QMS dispositions, or ERP writes through Permute. Any external action needs a supported provider/action scope, current permissions, and approval. Keep the signed study and authorized process change in the systems your quality procedure requires.

Process capability supports a qualification decision when the requirement, study population, and statistical assumptions hold together. Return to the oversized shaft results with the applicable drawing revision and evidence of process behavior over time. Use Cp to examine spread and Cpk to examine the nearer specification margin under the declared method. Preserve the acceptance decision and its supporting records so the next weekly review can compare the same process conditions.

Review the gaps in your quality data

Bring an inspection extract, its drawing history, and production references to assess what a reproducible study needs.

Questions about qualifying a manufacturing process

Can a pass/fail inspection export support Cp and Cpk?

These indices use the spread and center of continuous measurements. A pass/fail flag discards the measured distance to the limit, so retrieve the underlying measurements or use a method suited to attribute data. Preserve the acceptance counts for the separate yield review.

Can machines making the same part share one study?

Check the measurements by machine before pooling them. A common part number does not establish that machines have the same center or variation. Keep separate studies when their process behavior differs, and document the scope if the quality engineer approves a combined analysis.

Can an in-control process still fail specifications?

Yes. Stability describes process behavior over time; the specification describes the required result. A stable process can have a mean too close to a limit or too much variation, which is why the capability review follows the stability review.