Process Capability Index Calculator

Capability indices compress a process and its specification into one question: does the distribution fit inside the tolerance, with room to spare? Enter the spec limits, the process mean and sigma, and this calculator returns Cp (the potential), Cpu and Cpl (each side’s margin), and Cpk (the reality), plus the defect rate they imply.

Potential Versus Reality: Cp and Cpk

The pair exists to separate two failures that look identical on a shipping dock. Cp compares the tolerance width to the process's natural spread (6σ): it asks only whether the distribution could fit, and it is blind to where the process is aimed. Cpk takes the worse of the two one-sided margins — distance from the mean to the nearer spec limit, in units of 3σ — and so prices aim and spread together. A Cp of 1.67 with a Cpk of 0.8 is a precise process pointed at the wrong place (recentre it and the defects vanish); a Cp of 0.8 is a process too wide for its tolerance however it is aimed (only variation reduction, or a wider spec, saves it). The calculator prints both, plus the two one-sided indices that identify the guilty wall, because the corrective actions for the two failures share no overlap and confusing them is expensive.

Worked Example: A Well-Behaved Machining Process

The defaults describe a turned part: specification 9.8 to 10.2 (a ±0.2 tolerance on a nominal 10), process mean 10.01, sigma 0.04. Cp = 0.4 ÷ (6 × 0.04) = 1.67. The upper margin: Cpu = (10.2 − 10.01) ÷ (3 × 0.04) = 1.58; the lower: Cpl = (10.01 − 9.8) ÷ 0.12 = 1.75. Cpk = min = 1.58 — capable by the usual 1.33 bar, nearly at the 1.67 tier, with the upper wall the nearer one, exactly as a mean sitting a hair above nominal implies. The implied defect rate, summing the two normal tails at 4.75σ and 5.25σ, lands near 1 ppm — the calculator prints the figure from the distribution rather than from the index, since the same Cpk can hide different tail splits when the process sits off-centre. One-ten-thousandth of the tolerance budget is being spent on aim error; the rest is spread. That is a healthy process's portrait.

The Sigma and the Stability It Must Earn

Every figure on this page is only as real as the sigma fed into it, and sigma has a provenance the formula cannot check. The convention (AIAG's manuals are the reference dialect) is that capability sigma is the within-subgroup, short-term variation — estimated from control-chart ranges (R̄/d₂) rather than the raw standard deviation of a mixed pile of parts — because it represents the process's inherent voice with assignable-cause noise removed. Feed in a long-term sigma instead and you have computed the Pp/Ppk family, the performance indices, which answer the neighbouring question of what the process did rather than what it can do. And beneath both sits the assumption the note repeats: the process must be in statistical control. An unstable process has no single distribution, hence no capability — only a history. Control first (the hub's control-limit calculator is the upstream tool), then capability, then celebration, in that order.

One-Sided Specifications: Half the Indices Retire

Many real characteristics carry a single wall — a maximum impurity, a minimum strength, an upper limit on roughness — and the capability family adapts without drama: only the relevant one-sided index exists (Cpu for an upper-only spec, Cpl for a lower-only), Cp is undefined (there is no width to fit into), and Cpk simply is the surviving one-sided index. This calculator's form expects two limits because the two-sided case is the grammar the family is taught in, but the one-sided reading is immediate from its outputs: quote the side that exists and let the absent wall go. What does not adapt is the discipline underneath — stability first, an honest sigma, and a distribution whose tail toward the single wall has actually been looked at, since with one wall there is no other side to average the risk away.

Capability Is Not a Grade, It Is a Risk Statement

Read the index as what it actuarially is: the ppm line converts it into the expected number of defective parts per million opportunities, and that figure — not the tidy two-decimal index — is what a customer experiences as warranty claims and a plant experiences as sorting cost. The translation disciplines the thresholds of the FAQ: the step from Cpk 1.00 to 1.33 is not a third more virtue, it is a forty-fold cut in worst-case defect escape (about 1,350 ppm to about 33 ppm on the governing side), and the step to 1.67 drives the governing tail below one part per million — another hundred-fold beyond that. This is why capability requirements sit in contracts with decimal places that look fussy until the ppm column is priced: each 0.33 of index is an order of magnitude of risk, and the calculator prints both dialects so the negotiation can happen in either.

When the Index Improves Without the Process Changing

A cautionary tale the quality literature repeats because factories keep re-enacting it: Cpk can be raised three ways that ship the same defects — widen the specification by negotiation (sometimes legitimate, always visible), measure less precisely so the tails smear politely (fraud with a calibration certificate), or study a friendlier slice of production than the customer receives (the golden-morning capability study). The index cannot police its own inputs; the governance around it — frozen specs, calibrated gauges with their own R and R studies, random sampling across the process's whole life — is what makes the number in this calculator mean the same thing in the customer's plant as in yours. Capability is a measurement system with a formula at its centre, and the formula is the easy part, which this page happily does for you.

Related Tools

More from the Statistics Calculators hub: the Upper Control Limit Calculator — the stability check that must precede this page — and the DPMO Calculator for the defect dialect of the same story.

Frequently Asked Questions

What Cpk value is acceptable?

The industry's working thresholds: 1.00 is the bare floor (the nearer wall at 3σ — about 1,350 ppm escaping on that side alone, 2,700 ppm in total when centred), 1.33 is the common production requirement (4σ to the wall, roughly 63 ppm in total when centred), and 1.67 is demanded for safety-critical and new-process approvals (5σ). Customer contracts routinely specify which, and automotive supply chains enforce them through PPAP. The calculator prints the verdict bands against these conventions while keeping the number itself in view — thresholds are policy; the index is measurement.

Can Cpk be negative?

Yes, and it is the index's most honest moment: if the process mean sits outside the specification, the nearer margin is negative and Cpk follows it below zero. More than half the output is scrap by construction, and the negative index says so in one signed figure where Cp would still print the same sunny potential as ever. Any capability report that cannot go negative is hiding the worst case a process can be in.

What is the difference between Cp/Cpk and Pp/Ppk?

The sigma. C-indices use within-subgroup (short-term) sigma and describe capability — what the process can do when stable. P-indices use the overall (long-term) standard deviation of all the data and describe performance — what it actually did over the period, drift, shifts and all. Ppk at or below Cpk is the normal relationship; a large gap between them is the signature of a process whose centre wanders, which is a control problem wearing a capability costume.

How does Cpk convert to a defect rate or sigma level?

Through the normal tails, as the calculator computes: the nearer wall's z is 3 × Cpk, the farther wall's z comes from the other one-sided index, and ppm is the summed tail probability. The "sigma level" of Six Sigma fame adds the conventional 1.5σ long-term shift to that z — a convention about drift, not a property of your data — so a Cpk of 1.5 (4.5σ to the wall) corresponds to the famous 3.4 ppm only after that shift is applied to one side. This page reports the unshifted, measured figure and names the convention when it matters.

My data is not normal. Do the indices still work?

The indices as pure ratios (widths in sigma units) can be computed for anything; the ppm translation is where normality is spent. Skewed processes — tool-wear ramps, one-sided characteristics like flatness or impurity — need either a transformation (Box-Cox, Johnson), a percentile-based capability method, or a distribution fitted to the data, and quoting the normal-theory ppm for them can be wrong by orders of magnitude. The calculator's ppm line carries its assumption in its first clause; where the histogram disagrees with the bell, believe the histogram.

How many parts should I measure for a capability study?

Enough that the sigma estimate itself is stable — the manuals' working figures are at least 25 subgroups or 100 individual measurements for a formal study, taken across the process's ordinary range of operators, material lots and setups (a study run on one golden morning measures the morning). Fewer data can still orient a conversation, but the index inherits the sample's uncertainty, and a Cpk quoted to two decimals from fifteen parts is precision theatre. Measure representatively, then compute.