Control-Limit Calculator: Set Limits That Hold the Gain After the Team Moves On

The Control-Limit Calculator computes upper and lower control limits at three sigma for each monitored metric, using baseline process statistics. You give it a centre and a sigma for every metric. It returns the limits and tells you whether the limits already written into your control plan agree with the arithmetic.

That sounds small. It is not. A control plan lives or dies on its thresholds. Set them too wide and the chart never signals, so the improvement decays unnoticed. Set them too narrow and the process owner chases noise, loses faith in the chart, and stops looking at it. Either way the gain leaks away quietly in the months after the project team disbands.

Use the tool in the Control phase of RDMAICS, once your improvements are stable and you have fresh baseline statistics from the improved process. Use it again at handover, when the new process owner signs for the plan. Use it a third time at the Sustain review, when enough post-project data exists to test whether the original limits still describe reality.

The tool ships with sample data loaded, and this article is built entirely around that sample. The case is fictional. Treat it as a teaching example, not a benchmark.

The sample describes a control plan with four monitored metrics drawn from a single transactional process. Field error rate sits on a p chart with a centre of 0.08 and a sigma of 0.024. Cycle time sits on an I-MR chart with a centre of 3.9 and a sigma of 0.6. Batch error count sits on a c chart with a centre of 5 and a sigma of 2.2. Defects per unit sits on a u chart with a centre of 0.75 and a sigma of 0.18.

Four metrics is a realistic control plan size. Enough to cover the failure modes the project addressed. Few enough that someone will actually review them each week. The mix of chart types is also realistic, because a single process rarely produces a single kind of data.

Each row in the table carries the same six things. The metric name. The centre, which is a mean for continuous data and a p-bar for proportion data. The sigma or standard deviation. The chart type. The upper and lower control limits you currently have written down. A notes field for anything that explains how the sigma was derived.

The notes column in the sample does real work. Field error rate carries the note "Variable n". Cycle time carries "n=1, MR-bar/d2". Batch error count carries "Constant area". Defects per unit carries "Variable n". Those notes are the audit trail. Six months from now, someone will ask where the sigma of 0.6 came from, and the answer is sitting in the row.

Enter your existing UCL and LCL even when you suspect they are wrong. The tool compares what you entered against what the arithmetic produces. Leaving the field blank removes the comparison, which is the most useful part of the output. In the sample, the entered pairs are 0.15 and 0.01 for field error rate, 5.7 and 2.1 for cycle time, 12 and 0 for batch error count, and 1.3 and 0.2 for defects per unit.

The output opens with a short statement of what was done: control limits recalculated as centre plus or minus three sigma. Beneath that sit four counts. Four metrics. Four limits recalculated. Entered limits agree, four of four. Lower limit set to zero, one.

Then come the per-metric results. Field error rate returns an LCL of 0.008, a centre of 0.08 and a UCL of 0.152. Cycle time returns 2.1, 3.9 and 5.7. Batch error count returns 0, 5 and 11.6. Defects per unit returns 0.21, 0.75 and 1.29.

The agreement line is the one to read first. The entered UCL and LCL agree with centre plus or minus three sigma for all four metrics, within five per cent of the limit span. That is a clean result. It means the control plan was built with the same arithmetic the tool uses, and nobody has edited a limit by hand since.

The zero count matters too. One metric had its lower limit set to zero. That is batch error count on a c chart, where centre minus three sigma falls below zero and a count cannot be negative. The tool truncates it rather than printing a meaningless negative. A short AI review is available alongside the figures if you want the reasoning explained in prose.

The first decision is whether to publish the control plan as written. When all entered limits agree with the recalculation, as they do across all four metrics in the sample, you publish. When they disagree, you find out why before anyone signs.

The second decision is what the process owner does with an out-of-limit point. A limit is only actionable if someone knows what happens when it is breached. The calculated figures give you the trigger values to write into the reaction plan. Cycle time above 5.7 or below 2.1 means something changed. Defects per unit above 1.29 means something changed.

The third decision is about review cadence. Metrics with tight limits relative to their centre will signal more often. Field error rate runs from 0.008 to 0.152 around a centre of 0.08, which is a wide relative span. Cycle time runs from 2.1 to 5.7 around 3.9. You can use the spans to judge how much attention each chart will demand before you hand it over.

The fourth decision is handover readiness. A control plan where the arithmetic has been checked, the chart types are declared and the sampling basis is recorded is a plan a new owner can defend. One where limits were typed in from memory is not.

The most common mistake is reusing pre-improvement sigma. If the project reduced variation, the old sigma produces limits that are too wide, and the chart will never signal. Recompute sigma from the improved process before you enter anything.

The second is mislabelling the chart type. A p chart and a u chart behave differently at the lower boundary and under variable sample size. The sample marks two metrics as having variable n. That note is a reminder that the limits shown are an approximation at an average sample size, not a fixed line for every point.

The third is allowing a negative lower limit to stand. Counts and proportions cannot go below zero. The tool handled this once in the sample, on the c chart. If your own output shows a negative lower limit after the chart type is declared correctly, check the chart type again.

The fourth is treating agreement as proof of correctness. The tool confirmed that the entered limits match centre plus or minus three sigma for all four metrics. That confirms internal consistency. It does not confirm that the centre and sigma were measured well.

Control limits are the mechanism that converts a finished project into a standing process. Without them, the improved process has no definition of normal, and drift has nothing to trip against.

Print the calculated limits into the control plan document. Attach the notes column. Name the owner for each metric. Write the reaction plan against the specific numbers, not against the phrase "out of control". A reaction plan that says "if cycle time exceeds 5.7, escalate" will be followed. One that says "monitor for special cause" will not.

Then set a date to recalculate. The limits in the sample describe the process as it was when the baseline was taken. Processes move. Volumes change. The Sustain phase exists to catch that, and this tool is how you check it.

It will not tell you whether your process is capable. Control limits describe the voice of the process. Specification limits describe the voice of the customer. A metric can sit comfortably inside its control limits and still fail every customer requirement. The tool says nothing about that gap.

It will not validate your centre or your sigma. It takes the numbers you give it and does arithmetic on them. If the sigma of 0.6 for cycle time was estimated from a short, unrepresentative sample, the limits of 2.1 and 5.7 will be wrong and the tool will still report agreement.

It will not test for stability. Three sigma limits assume the baseline data came from a stable process. If your baseline period contained a shift, the sigma is inflated and the limits are too wide. Check stability before you compute limits, not after.

It will not handle variable sample size point by point. Two of the four sample metrics are marked "Variable n". For those, the real limits step up and down with each sample. The figures here are a single representative pair. Use them for the control plan document and let your charting method handle the per-point variation.

Finally, it will not make anyone look at the chart. That is a management problem, not a calculation problem. The tool gives you defensible numbers. The rest is discipline.