If you have ever written or approved a packaging spec like “My box is 10.000 inches +/- 1/8 inch,” it feels clear. But it hides the real question: What does that tolerance mean when you are making 1,000,000 boxes?
How many will be out of spec? How many will jam your line? And can your supplier’s process even hold what you just put on the drawing?
This article is a practical, packaging-focused walkthrough of three ideas that every packaging engineer should understand:
- Cpk / process capability: can the process hold the tolerance?
- AQL acceptance sampling: how many defects are allowed before rejecting a lot?
- Tolerance analysis / interference fit: what happens when product and package variation collide?
1. The “10 Inches +/- 1/8” Problem: What It Means at Scale
Consider a typical corrugated box dimension spec:
- Nominal = 10.000 in
- Tolerance = +/- 0.125 in
- LSL (Lower Spec Limit) = 9.875 in
- USL (Upper Spec Limit) = 10.125 in
- Total spec width = 0.250 in
That spec does not tell you:
- How tight the supplier’s process actually is.
- Whether the process is centered on the target.
- How many out-of-spec units you will see over time.
Two suppliers can both “target 10 inches” and have very different outcomes:
- Supplier A: Tight process, well-centered, almost no defects.
- Supplier B: Wider process, slightly off-center, thousands of defects per million.
This is exactly what process capability (Cpk) quantifies.
2. Cpk: “Can the Machine Actually Hold the Tolerance?”
Every manufacturing process has natural variation. Process capability compares the allowed spread (your spec limits) to the actual spread (process standard deviation). NIST describes process capability as comparing an in-control process output to spec limits using capability indices (NIST Engineering Statistics Handbook).
The formula:
Cpk = min(Cpu, Cpl)
where:
Cpu = (USL - mean) / (3 * sigma)
Cpl = (mean - LSL) / (3 * sigma)Why “min()” matters: Being off-center kills you. A process can have small variation but still produce defects if the mean drifts toward one spec limit. Cpk captures the worst side of the distribution relative to the nearest limit.
Practical threshold
A Cpk of 1.33 or higher is widely regarded as the practical minimum for a capable process. Below 1.33, defect rates rise quickly and process drift becomes a real production risk. Many packaging operations target 1.67+ for critical dimensions on high-speed automated lines.
3. A Concrete Packaging Example: 1,000,000 Boxes
Using the spec from Section 1, suppose the supplier’s measured process data shows:
- Mean = 10.000 in (perfectly centered)
- Standard deviation (sigma) = 0.040 in
Computing Cpk:
Cpu = (10.125 - 10.000) / (3 * 0.040) = 0.125 / 0.120 = 1.04
Cpl = (10.000 - 9.875) / (3 * 0.040) = 0.125 / 0.120 = 1.04
Cpk = min(1.04, 1.04) = 1.04Interpretation: A Cpk of 1.04 means the process is barely capable. With 1,000,000 units, the out-of-spec count becomes very real, roughly 1,350 defective parts per million on each side of the distribution.
Now watch what happens if the process mean drifts by only 0.040 in (one sigma):
New mean = 10.040 in
Cpu = (10.125 - 10.040) / 0.120 = 0.085 / 0.120 = 0.71
Cpl = (10.040 - 9.875) / 0.120 = 0.165 / 0.120 = 1.38
Cpk = min(0.71, 1.38) = 0.71A Cpk of 0.71 is a failing process. That is the “million boxes” reality: tiny mean shifts create enormous defect counts.
You can run these calculations directly using the Statistics Suite, which includes process capability analysis with Cpk, Cp, and sigma-level outputs.
4. What Cpk Number Should You Require for Packaging?
| Cpk Threshold | Interpretation | Typical Use |
|---|---|---|
| >= 1.00 | ”Capable-ish”: expect issues over time, especially with process drift | Low-volume, non-critical dimensions |
| >= 1.33 | Widely used practical capability threshold | Most packaging dimensions, general manufacturing |
| >= 1.67+ | High capability: failures are rare even with drift | High-volume automated lines, medical/pharma, regulated industries |
Your “right answer” depends on four factors:
- Volume: Higher volume amplifies even small defect rates.
- Cost of a defect: A jammed line or a field return is far more expensive than a scrapped blank.
- Interaction: Does this dimension mate with another part? (See Section 7 on interference fit.)
- Measurement system quality: Your Cpk number is only as good as your gauge.
Gauge R&R affects capability results
If your measurement system has high variation (poor gauge R&R), it inflates the observed sigma and understates your true Cpk. Always validate your measurement system before drawing conclusions from capability data. A gauge R&R study should show measurement variation consuming less than 10% of the total tolerance to be considered adequate. The Statistics Suite includes a dedicated GR&R analysis mode for this purpose.
5. AQL in Packaging: “How Many Bad Boxes Before Rejecting the Lot?”
AQL (Acceptable Quality Level / Acceptable Quality Limit) is used in acceptance sampling. The primary reference standards are:
- ISO 2859-1: an attribute sampling system indexed by AQL (ISO).
- ANSI/ASQ Z1.4: the U.S.-aligned equivalent.
AQL does not mean “We allow 1% defects.” It is a parameter used to design sampling plans with known statistical behavior. Specifically, it defines the tradeoff between producer risk (rejecting a good lot) and consumer risk (accepting a bad lot).
AQL vs. Cpk: different questions, complementary answers
Cpk answers: “Is the process capable of meeting the spec consistently over time?” It is a process-level metric. AQL sampling answers: “Given we cannot inspect everything, what sample-and-decision rule do we use for this specific lot?” It is a lot-level inspection tool. If you rely on AQL instead of capability, you are saying: “We expect defects; we will just try to catch enough of them.” That is how high-volume pain starts. Use both together.
Packaging defect categories commonly used with AQL sampling:
- Critical: Wrong IFU, seal breach, contamination risk. AQL typically 0.065 or 0.10.
- Major: Dimension prevents assembly, carton will not erect, structural failure. AQL typically 1.0 or 2.5.
- Minor: Cosmetic scuff, slight color shift, non-functional blemish. AQL typically 4.0 or 6.5.
PackCalc’s Statistics Suite includes a sampling plan analysis mode where you can input lot size, AQL level, and inspection level to generate accept/reject numbers aligned with ISO 2859-1 / ANSI Z1.4 tables.
6. Where AQL and Cpk Get Mixed Up
Teams often confuse these two tools because both involve “defects” and “quality.” Here is the distinction:
- Cpk is about the process: Is it stable, centered, and tight enough to consistently produce conforming output?
- AQL sampling is about the inspection decision: Given a finite lot, how do we decide to accept or reject it?
These are complementary, not interchangeable:
- Use Cpk to qualify the process and set specifications that the supplier can actually hold.
- Use AQL sampling as an ongoing risk-based inspection check to catch lots that slip.
If you skip capability and rely only on AQL sampling, you are accepting that defects will happen and hoping your sampling catches enough of them. For high-volume packaging lines, this is a losing strategy.
7. Interference Fit in Packaging: When Variation Makes the Product “Bigger Than the Box”
Product dimensions and packaging dimensions both vary. Even if both drawings look “fine” in isolation, overlap in the distributions creates insertion force spikes, jams, cracked trays, and field returns.
Worst-case stack analysis (conservative):
- Max product size + Min package opening = worst interference.
- Simple and safe, but often over-designs the package.
Statistical stack analysis (realistic):
- Treat both product and package dimensions as distributions.
- Combine using RSS (Root Sum of Squares) or Monte Carlo simulation.
- Compute the probability of interference at volume.
This is where Cpk becomes directly actionable: if both the product and the package have Cpk >= 1.33, the probability of interference drops dramatically compared to two processes running at Cpk = 1.00.
8. The Hidden Trap: Tolerances Without Capability Are Just Wishes
Teams set tolerances by:
- Copying old drawings.
- Eyeballing clearances.
- Picking “neat” fractions (1/8, 1/16).
- Over-trusting supplier claims.
None of these approaches considers what the process can actually deliver. NIST explicitly frames capability in terms of an in-control process compared to specification limits. The spec is only meaningful if the process can hold it (NIST Engineering Statistics Handbook).
The real workflow:
- Stabilize the process: get it in statistical control (SPC charts).
- Measure variation correctly: validate the gauge (R&R study), collect enough data.
- Compute Cp/Cpk: quantify whether the process can hold the tolerance.
- Set AQL sampling as a risk-based ongoing check, not a substitute for capability.
Use the Statistics Suite before locking drawings
Run your dimension data through the Statistics Suite to compute Cpk, estimate defect rates at volume, and validate tolerance scenarios before you freeze the specification. It is far cheaper to widen a tolerance on a drawing than to reject truckloads of material that a supplier’s process cannot hold.
A) Glossary (short)
- Cpk (Process Capability Index): A statistical index that measures how well a process output fits within specification limits, accounting for both spread and centering. Cpk = min((USL - mean) / 3sigma, (mean - LSL) / 3sigma).
- Cp (Process Capability): Similar to Cpk but assumes the process is perfectly centered. Cp = (USL - LSL) / (6sigma). Useful for understanding potential capability.
- AQL (Acceptable Quality Limit): A parameter used in acceptance sampling plans (ISO 2859-1, ANSI/ASQ Z1.4) to define the quality level that is the worst tolerable process average when a continuing series of lots is submitted for inspection.
- ISO 2859-1: The international standard for sampling procedures for inspection by attributes, indexed by AQL.
- ANSI/ASQ Z1.4: The U.S.-aligned equivalent of ISO 2859-1, widely used in packaging and manufacturing acceptance sampling.
- Gauge R&R (Repeatability and Reproducibility): A study that quantifies how much of the observed measurement variation is due to the measurement system itself versus the actual part-to-part variation.
- SPC (Statistical Process Control): The use of control charts and statistical methods to monitor and control a process, ensuring it operates at its full potential.
- RSS (Root Sum of Squares): A statistical method for combining independent tolerances, yielding a more realistic (less conservative) stack-up than worst-case addition.
- Interference Fit: A condition where the product dimension exceeds the package opening dimension, causing insertion difficulty or damage.
- LSL / USL: Lower Specification Limit and Upper Specification Limit, the boundaries of acceptable variation defined on a drawing or spec.
Citations included from NIST Engineering Statistics Handbook, ISO, and ASTM International as noted in text.