Operations & Systems

Run-Sequence Plots for Business Processes: Keep the Order Before Explaining the Change

The report contains every duration. It has lost the order in which the work happened.

Someone sorts the column from shortest to longest. The chart rises neatly. The team starts explaining an upward trend that the sorting operation created.

Keep the sequence before explaining the change.

A run-sequence plot places a measured value against its observation index. NIST's run-sequence guidance uses that view to inspect changes in location and variation. Its vertical axis carries the response; its horizontal axis carries the sequence number.

The operating method below applies that idea to a small service report. The worksheet, records, and process-change story are invented. They establish no business result, statistical stability, or cause.

Define One Observation Before Drawing the Series

Start with a sentence another person can reproduce: one observation is one eligible standard request's processing duration, measured in elapsed minutes.

For this fictional report, processing begins when review starts. It ends when the reviewer records the first completed response. Waiting before review is outside the measure. Time spent waiting during review remains inside it. Reopened requests are outside this particular series.

Those choices define this example. A different service can need different boundaries. Write yours in a business metric dictionary before comparing reports with similar labels.

Record the collection rule too. This example includes twelve consecutive eligible requests. It orders them by review-start timestamp, with a source event identifier resolving timestamp ties. It does not order them by completion, duration, or reviewer.

That distinction matters even inside an invented table. A request that starts earlier can finish later. Calling completion order “the original order” would leave the reader guessing which sequence you retained.

Keep the source extract, eligibility rule, timestamp precision, time basis, and ordering keys with the report. If tied timestamps have no defensible secondary key, mark their order uncertain. Do not silently manufacture precision through row numbers.

Also state which conditions should remain comparable. Here, the request type, timing definition, and collection method stay constant. A known instruction change appears in the record. It belongs in the interpretation, without becoming an explanation by default.

Read the Complete Fictional Duration Table

Each row below represents one request. Dates are synthetic labels in the same reporting time basis. Two eligible requests begin on each listed date. The observation index follows their recorded start order.

Index Request Review-start date Duration in minutes Recorded context
1 R-01 2026.10.01 18 Earlier instruction
2 R-02 2026.10.01 21 Earlier instruction
3 R-03 2026.10.02 19 Earlier instruction
4 R-04 2026.10.02 20 Earlier instruction
5 R-05 2026.10.03 22 Earlier instruction
6 R-06 2026.10.03 20 Earlier instruction
7 R-07 2026.10.04 30 Revised instruction recorded
8 R-08 2026.10.04 33 Revised instruction recorded
9 R-09 2026.10.05 29 Revised instruction recorded
10 R-10 2026.10.05 34 Revised instruction recorded
11 R-11 2026.10.06 31 Revised instruction recorded
12 R-12 2026.10.06 35 Revised instruction recorded

Read the first six observations as a group, then the next six. The earlier values sit between eighteen and twenty-two minutes. The later values sit between twenty-nine and thirty-five minutes.

The first group totals one hundred twenty minutes across six requests, giving a mean of twenty minutes. The second totals one hundred ninety-two minutes, giving a mean of thirty-two minutes. All twelve total three hundred twelve minutes, with a mean of twenty-six minutes.

These are arithmetic descriptions of invented records. The combined mean preserves the total but removes the separation visible in sequence. Neither group mean estimates an established process parameter here.

The later group also has a larger observed range: six minutes instead of four. That small comparison suggests another question about spread. It does not establish a lasting change in process variation.

Compare Sequence Order With Size Order

The following table presents the same twelve durations twice. The left value retains the original observation index. The right value occupies a rank after sorting by size.

Position Original sequence, minutes Sorted by size, minutes
1 18 18
2 21 19
3 19 20
4 20 20
5 22 21
6 20 22
7 30 29
8 33 30
9 29 31
10 34 33
11 31 34
12 35 35
The same fictional durations in observation order and size order Both panels use a zero to forty minute vertical scale. The upper panel preserves observation indices one through twelve and marks the instruction change between six and seven. The lower panel sorts the same values from smallest to largest. Every value appears in the adjacent table. Original observation order Instruction change: before observation 7 40200 1612 Observation index Values sorted by size 40200 1612 Size rank MinutesMinutes
Two views of the same twelve synthetic values. The upper plot retains the recorded change boundary. The lower plot uses rank. The table supplies every value without requiring visual interpretation.

Sorting leaves the sum and mean unchanged. It removes the connection between each value and its original sequence position. In the right column, position seven means seventh smallest. It no longer means the seventh request observed.

For a sequence plot, use the left column with index on the horizontal axis and minutes on the vertical axis. Label those axes explicitly. Place the instruction-change annotation between observations six and seven.

The right column can describe the ordered set of sizes. If you graph it, label the horizontal axis as size rank. Its upward arrangement cannot serve as evidence that processing durations increased through the original sequence.

In this deliberately simple sample, both columns still reveal a gap between twenty-two and twenty-nine minutes. The original sequence adds information: all six larger values occur after all six smaller ones. A size ranking alone cannot establish that relationship to the recorded instruction change.

Keep request identifiers available when you inspect either view. A reader who selects thirty-three minutes should be able to find R-08. A chart becomes harder to investigate when it disconnects the mark from its source record.

Annotate the Known Change Without Assigning Cause

The fictional change record says the revised instruction takes effect before R-07 starts on 2026.10.04. Annotate that boundary as context. Keep the exact source note with the report.

The visible finding is narrow: every later observation exceeds every earlier observation in this sample. The instruction change coincides with that division. The example contains no evidence that isolates its effect.

Ask what else differs across the boundary. Request complexity might differ. A timing export might change. An instruction might add a required review step. These are candidate explanations to inspect, not facts supplied by the duration column.

Choose evidence that can distinguish those candidates. Compare the request categories. Inspect timing event definitions. Read the effective instruction versions. Check whether the eligibility filter changes at the same point.

Use process variant analysis when the question concerns different recorded paths through the work. A longer duration does not by itself identify an added approval, repeated review, or missing activity.

Keep the next question bounded: do R-07 through R-12 use the same start and end events as R-01 through R-06? Answering that question tests comparability before anyone proposes a remedy.

Keep Missing Records and Definition Changes Visible

Suppose a later source check reveals an eligible request between R-06 and R-07 whose duration is unavailable. That request is absent from the original fictional extract. The earlier claim of twelve consecutive eligible requests now needs revision.

Retain the original report version. Record the newly identified request at its known sequence position with duration unknown. Do not enter zero. Do not claim that a connecting segment describes the missing duration.

If the missing request's position is also unknown, say so. Keep that uncertainty in the series boundary. A tidy chart should not decide an ordering question that the source cannot answer.

Now consider a different hypothetical discovery. The newer export measures from request receipt, while the older export measures from review start. These values represent different intervals. A date annotation alone does not make them comparable.

Separate those definitions until you can reconstruct a consistent measure from the underlying events. Preserve both original labels. Document any reconstruction rule and retain the records it uses.

Also distinguish observation index from elapsed calendar time. Two adjacent points can represent requests separated by minutes or days. If the decision depends on calendar spacing, provide timestamps or a separate time-based view. Do not make equal index spacing stand for equal elapsed time.

Use a Plot-Reading Worksheet

Copy this proposed worksheet into the report's working notes. Use one row for each claim that requires a different source check.

Series boundary Observed pattern Alternative explanation Source evidence needed Next question
Twelve eligible requests, ordered by review start Later six durations exceed earlier six Different request mix Categories and eligibility records Are request types comparable across the boundary?
Same twelve records Instruction change aligns with later group Timing definition also changed Event definitions and export configuration Do both groups measure the same interval?
First six versus last six Observed ranges are four and six minutes Short sample happens to differ Longer comparable ordered history Does the wider spread persist in additional observations?
Revised extract after missing-record discovery One duration is unknown Incomplete extraction Source request and export log Is the record missing from the export or missing its end event?

The worksheet separates what the values display from what an investigator proposes. Leave an explanation unresolved until the requested evidence arrives. The chart does not fill that field for you.

Give each next question an owner and a review point in your working record. Keep the table readable for someone who cannot use the figure. Include units and the complete observation list beside any plotted version.

For an operator dashboard, carry the bounded finding into the review note. Keep the source table and worksheet available behind it. A short dashboard label should not erase the definition or unresolved explanation.

Close With the Question the Evidence Supports

NIST's guidance treats the run-sequence view as a way to inspect assumptions about location and scale. It does not supply formal control-chart limits or a causal test for this fictional service example.

Write the conclusion at the level you checked: these twelve records show two separated groups in their retained order. A recorded instruction change sits at their boundary. Request mix, measurement consistency, and missing history still need review.

Avoid declaring the process stable from a short, quiet-looking segment. Avoid assigning an unusual point to an employee without inspecting the source conditions. Keep formal signal rules outside this exploratory reading unless a qualified method establishes them.

The useful next action is a source check with an answerable question. Preserve the order that makes that question visible.

Sources

Topic map: Business reporting and statistics

Articles in this cluster, then the 3 clusters that share the most topics with it.

This cluster: Business reporting and statistics

Neighbor: Spreadsheets, CSV files, and data handoffs

Neighbor: Records, access, and continuity

Neighbor: APIs, webhooks, and automation