Control limits come from your process, calculated from its own data, usually three standard deviations from the mean, and show what it naturally does. Specification limits come from the customer and define what is acceptable. They answer different questions and never belong on the same chart.
Confusing these two is one of the most common and most costly mistakes in statistical process control. Put spec limits on a control chart and you will chase the wrong signals, adjust a stable process, and miss a drifting one. Keep them straight and each does its job: control limits tell you whether the process is stable, spec limits tell you whether the parts are good, and the gap between them is where capability lives.
What is the difference between control limits and specification limits?
Control limits describe the process; specification limits describe the requirement. Control limits are calculated from the process's own output, you collect data, find the average and the natural variation, and the limits fall out at roughly three standard deviations either side of the mean. They are not a target and not a wish; they are a measurement of what the process actually does when it is running normally. Move to a more consistent process and the control limits tighten on their own, because the data changed.
Specification limits are the opposite in origin. They come from outside the process entirely, the customer's drawing, a regulation, a fit requirement with a mating part. A shaft that must be 10.00 mm plus or minus 0.05 has spec limits of 9.95 and 10.05 no matter how the process behaves. You cannot calculate a spec limit from your data, and you cannot make a bad spec good by running a tight process. The two limits are set by different people, for different reasons, and they can sit anywhere relative to each other.
Where does each come from?
Control limits are the voice of the process; specification limits are the voice of the customer. That pairing is the fastest way to keep them straight. The voice of the process is whatever the machine, material, and method produce together, you listen to it by plotting data on a control chart and letting the limits emerge. The voice of the customer is what someone downstream needs the part to be, handed to you as a tolerance before you ever cut metal.
Because they come from different places, neither one moves the other. Tightening your process does not change the customer's tolerance. Loosening the customer's tolerance does not change what your process does. This independence is the whole reason both exist: one tells you if you are stable, the other tells you if stable is good enough. A process can be dead stable and still miss the spec, and it can drift badly while every part still happens to pass, which is exactly why you need to read both, not one.
Why should you never put spec limits on a control chart?
Because a control chart is built to detect changes in the process, and spec limits carry no information about the process. Draw the customer's tolerance on your control chart and one of two bad things happens. If the spec is wider than your control limits, you will feel safe while the process shifts and drifts inside the spec, missing real signals that something changed, the chart has been blinded. If the spec is tighter than your control limits, points will cross it constantly and operators will "adjust" toward the spec line, which is tampering with a stable process and adds variation instead of removing it.
The control chart asks one question: has the process changed? It answers that by comparing points to limits derived from the process itself. Mixing in the customer's tolerance corrupts that comparison. Keep the chart clean, control limits only, and answer the separate question of "are the parts good?" with a capability study, where the spec limits belong.
Can a process be in control and still make bad parts?
Yes, and this is the single most important consequence of the distinction. A process is "in control" when it is stable and predictable, all points inside the control limits, scattering randomly. That says nothing about whether its output meets the spec. If the natural spread of a stable process is wider than the specification window, or if the process is centered off-target, it will produce defects every day while its control chart looks perfectly calm. Stability and acceptability are two different properties.
The classic way to see this is the four process states. A process can be in control and capable (the goal), in control but not capable (stable, but the spread does not fit the spec, so it makes predictable scrap), out of control but currently capable (drifting and unpredictable, passing today on luck), or out of control and not capable (chaos). A control chart alone only tells you which column you are in, stable or not. It cannot tell you which row, capable or not. For that you need a capability study.
How do control limits and spec limits connect through Cpk?
They connect through process capability, which is literally the ratio of the spec width to the process spread. Once a process is in statistical control, and only then, because capability math assumes a stable process, you can ask how well its natural variation fits inside the customer's tolerance. That is what Cpk measures: it compares the distance from the process mean to the nearer spec limit against three standard deviations of the process, so it accounts for both spread and centering.
A Cpk of 1.0 means the nearer spec limit sits about three standard deviations from the mean, the process spread just fills the spec, with essentially no margin. A common industry target is Cpk of 1.33 or higher, which puts a buffer between the process and the spec so ordinary variation does not produce defects. This is where control limits and spec limits finally meet: control limits (from the process) define the spread in the numerator's denominator, spec limits (from the customer) define the window, and Cpk is how they relate. You cannot compute it honestly until the control chart says the process is stable, which is why control comes before capability, always.
How do you use control limits and spec limits correctly?
Run them in the right order and keep them in their own lanes. The sequence matters because capability only means something on a stable process.
- Get the spec limits from the customer. Pull the tolerance from the drawing, standard, or contract. These are given, not calculated.
- Collect process data and plot a control chart. Measure the characteristic over time under normal conditions, enough subgroups to see the real variation.
- Calculate control limits from that data. Set them from the process mean and its natural variation (about three standard deviations out), using the right formula for your chart type. Never borrow the spec limits for this.
- Bring the process into control first. Investigate and remove special causes until the chart shows only random variation inside the control limits. Do not assess capability on an unstable process.
- Only now compare to the spec with a capability study. With a stable process, calculate Cpk against the spec limits to see whether the natural spread fits the customer's window and is centered.
- Act on the right lever. Out of control? Find the special cause. In control but not capable? That is a system problem, reduce common-cause variation or re-center, and a candidate for a corrective action.
The fundamentals, from the sources
- Control charts use limits calculated from the process data itself, conventionally at three standard deviations from the center line, to distinguish common-cause from special-cause variation (NIST/SEMATECH e-Handbook, control charts).
- Control limits reflect the actual behavior of the process and are not the same as specification limits, which state customer requirements; the two are set independently (ASQ, Control Chart).
- The NIST/SEMATECH Engineering Statistics Handbook Chapter 6 treats process control (stability) and process capability (meeting spec) as distinct techniques applied in sequence (NIST, Engineering Statistics Handbook, Chapter 6).
What are the common mistakes?
A few errors show up again and again, and every one of them traces back to blurring the two limits:
- Drawing spec limits on the control chart. It blinds the chart to process shifts or provokes tampering. Keep them separate.
- Assuming "in control" means "good." A stable process can make scrap all day if its spread is wider than the spec. Check capability, not just control.
- Calculating Cpk on an unstable process. Capability numbers from an out-of-control process are meaningless, because there is no single stable distribution to describe.
- Adjusting the process toward the spec line. Reacting to individual readings that approach the tolerance is tampering; it treats common-cause variation as if it were special cause.
- Widening control limits to stop alarms. Control limits are a measurement, not a dial. If the chart alarms too much, fix the process, do not move the limits.
All of this depends on actually having the data plotted and current. Control limits recalculated by hand once a quarter, or a capability study run from a stack of paper checksheets, are how these mistakes creep in unnoticed. Capturing measurements at the station and charting them live, the gap Harmony's quality intelligence is built to close, is what makes the difference between a control chart that catches a shift the same shift it happens and one that confirms a problem weeks after the scrap shipped. See our CLS case study for how that looks on a real floor.