Control limits and service targets answer different questions

Last updated: 5 Oct 20265 min read

tutorial

A control limit describes expected process variation under a frozen baseline, while a service target states an operational requirement.

Separate stability from adequacy

A support team may have a stable five-percent late-response rate while its contractual target is below three percent. Stability does not make the process acceptable. Conversely, a single late-response rate of 3.4 percent may miss the target without proving a change in the underlying process. Show both judgments: observed rate against the target, and observed rate against a process-monitoring rule. Mixing them creates false alarms or excuses a consistently poor service level. The metric denominator must be fixed first.

Choose a meaningful baseline

A process baseline should use periods with comparable case eligibility, clocks, staffing and logging. Do not blend a known outage week into the baseline merely because it adds more data. Keep the baseline date range and exclusion rationale in the release packet. A process chart assumes the baseline represents the system whose future variation is being monitored; if a new support policy changes that system, investigate and then deliberately establish a new baseline. Instrumentation changes can mimic process changes.

Interpret a signal as a prompt to investigate

A point outside a three-standard-deviation control limit suggests a special cause under the chart model. It does not identify the cause. Possible explanations include changed queue routing, a real backlog, a bad timestamp, and a denominator extract that lost eligible cases. A point within limits does not prove every case met the target or that the process has no slow drift. Alert investigation needs evidence beyond the chart.

Match chart design to the measurement

For a binary late/not-late outcome summarized by day, a proportion chart accounts for the number of cases each day. For a continuous duration, an individuals or subgroup chart needs a different variance model. Do not place a three-sigma band around daily percentages using the raw standard deviation of all individual case flags; the variability of a daily proportion depends on its denominator. The p-chart lesson makes that dependence explicit.

Make the decision rule reviewable

Store the target separately from the control center and limits. A day may be marked target miss, process signal, both or neither. This four-way state is more useful than one red/green badge. Confirm who investigates, what evidence they check, and whether an observed breach requires immediate action regardless of statistical signal. The project turns this into an auditable daily packet.

Implementation

def classify_service_day(late_cases, eligible_cases, target_rate, center, lower, upper):
    if eligible_cases <= 0 or not 0 <= late_cases <= eligible_cases:
        raise ValueError("invalid daily counts")
    if not 0 <= target_rate <= 1 or not 0 <= lower <= center <= upper <= 1:
        raise ValueError("invalid target or control limits")
    observed_rate = late_cases / eligible_cases
    return {"rate": observed_rate,
            "target_miss": observed_rate > target_rate,
            "process_signal": observed_rate < lower or observed_rate > upper}

steady_but_poor = classify_service_day(9, 180, 0.03, 0.05, 0.01, 0.09)
assert steady_but_poor["target_miss"]
assert not steady_but_poor["process_signal"]
unusual_but_on_target = classify_service_day(0, 180, 0.03, 0.05, 0.01, 0.09)
assert not unusual_but_on_target["target_miss"]
assert unusual_but_on_target["process_signal"]

Performance and operating cost

Classifying D daily summaries is O(D) time and O(D) output space if results are retained. The example takes limits as inputs; choosing and validating the chart model is separate work. A fast alert rule cannot repair an unstable metric definition.

Common Mistakes

Do not call an in-control process compliant with its service target.
Do not infer a root cause from one chart signal.
Do not recompute limits from the same day being judged.