Root-cause reasoning, improvement cycles and reproducible records

Preserve the observation and boundary

Define the defect, affected items, time window, operating states and measurement method. Compare the record with an appropriate reference. Include unaffected items and relevant conditions where they help discriminate explanations. A vague description such as “quality was worse” gives little basis for selecting a response or deciding whether it worked.

Draw the process and list candidate influences, including material, equipment, environment, methods and measurement. A cause-and-effect diagram can organise hypotheses and relationships. It is not a test result. Separate observed facts, plausible mechanisms, assumptions and missing information so that an attractive candidate does not silently become an established conclusion.

Seek evidence that could reject the explanation

In our original label case, one hypothesis predicts smearing follows material batch B across otherwise comparable setups; another predicts it follows setup S2 across comparable materials. A study that includes both batches and both setups can help distinguish the patterns. The allocation, repeat applications and measurement conditions need a suitable design.

Avoid changing several variables and then claiming the preferred one caused the improvement. If operational constraints require several changes together, report the package and the resulting limit on causal attribution. A response can restore the immediate service while leaving a separate investigation into mechanism and recurrence conditions. Keep those objectives and evidence distinct.

Let the designed result change the explanation

Complete the label example with original constructed results. On each of two days, prepare one fresh 50-label process run for each combination A/S1, B/S1, A/S2 and B/S2. Each run independently reapplies its specified setup to fresh material. One possible randomised order is day 1: B/S2, A/S1, B/S1, A/S2; day 2: A/S2, B/S1, B/S2, A/S1. Retain run IDs with day and combination. Hold the stated printer/ink configuration and inspection criterion fixed, log relevant conditions and code labels so the inspector does not know their combination. This is a teaching design and dataset, not a performed experiment or shop procedure.

The table keeps the two process applications in each cell visible. For S1, both batches have 3 smears among 100 inspected labels. Under S2, A has 4 among 100 and B has 15 among 100. The setup contrast S2−S1 is therefore 1 percentage point for A but 12 for B. Their difference, 12−1=11 percentage points, describes a material-by-setup interaction. Percentages are comparable because the denominator and defect definition are the same, not merely because a chart connects them.

The simple batch-only explanation predicted consistently higher smearing for B across setups; the simple setup-only explanation predicted a similar S2 increase for both materials. The supplied pattern instead points toward a combination-dependent explanation: S2 is associated with a much larger increase for B. This changes the investigation from declaring either batch B or S2 universally bad to asking why their combination differs. It does not yet identify the physical mechanism or prove that the observed contrasts persist in a population.

Two prepared runs per cell are the process replications; 100 inspected labels are not 100 independently reconfigured process runs. Chance variation, day effects, classification errors or an unrecorded condition could still matter. Preserve both days, inspect the run records and measurement agreement, and plan a confirmation under the same defined combinations with independent reapplications. A further material lot or operating condition addresses a different extension and should be identified as such. The retained comparison can support or challenge the explanation instead of becoming a pooled success claim.

Improvement includes learning after the change

An improvement cycle defines a problem, proposes a change, plans a comparison, examines results and decides what to retain or revise. Include relevant side effects and intended operating conditions. A lower observed defect fraction can reflect a different product mix, inspection rule or time window rather than the proposed improvement.

Preserve raw observations, item and run identifiers, allocation/order, versions, exclusions, analysis choices and uncertainty. Record why a conclusion follows and where it does not. Confirmation on a known configuration and monitoring over relevant conditions help distinguish an initial promising result from sustained behaviour. If later evidence disagrees, use the retained record to revise the explanation rather than deleting the inconvenient history.

Check denominators and the explanation

Suppose a constructed before period has 12 smears among 200 inspected labels, 6%. The after period has 6 among 100, also 6%. The count halved because the inspected total halved; the observed proportion did not improve. Even a lower after proportion would need a comparison that addresses variation and changed conditions before supporting a causal claim.

For the proposed batch/setup explanation, retain which combination produced each observation. A pooled total could hide that one combination improved while another worsened. Plan a confirmation whose conditions and outcomes could challenge the explanation, and document any revised process and measurement procedure so that the later evidence concerns the intended change.

For the separate designed comparison, pooling across setups gives A 7/200=3.5% and B 18/200=9%. That aggregate hides the contrast between equal 3% under S1 and 4% versus 15% under S2. Report the cell pattern and the two applications behind each cell. The discriminating result refines the explanation toward a combination effect; confirmation and measurement checks remain part of the next learning cycle.

Try the reasoning

1. Does placing a cause on a fishbone diagram establish it?

   Show the answer and reasoning

   No. The diagram organises possible influences; evidence is needed to distinguish and support them.

2. Why retain changes to an inspection rule?

   Show the answer and reasoning

   A different rule can change the observed defect count without changing the underlying product.

3. Did 12/200 to 6/100 improve the observed smear fraction?

   Show the answer and reasoning

   No. Both are 6%; comparing counts alone hid the changed denominator.

4. An unfamiliar comparison gives smears per 100: A/S1=2, B/S1=10, A/S2=3, B/S2=11. Which candidate pattern does it support descriptively, and what next check is useful?

   Show the answer and reasoning

   B−A is 8 percentage points under both setups, while S2−S1 is 1 point for both batches. The difference of setup contrasts is zero, so this supplied pattern has an additive batch difference rather than the earlier strong interaction. It favours a batch-associated explanation over an S2-only explanation in these records. Check independent repeat applications and coded inspection, then test relevant additional lots if generalisation is intended. One batch label does not identify a physical material property or prove causation.

5. Now suppose A/S1=3%, A/S2=11%, B/S1=10%, B/S2=4%, each with the same denominator and comparable run design. Why would pooled batch totals obscure the useful result?

   Show the answer and reasoning

   The average is 7% for both A and B, but B−A is +7 percentage points under S1 and −7 under S2. The batch contrast changes sign; setup contrasts are +8 for A and −6 for B, differing by −14 points. Preserve the cells and repeat a comparison capable of checking that crossover, including measurement and run-order records. Equal pooled means do not establish that changing setup or material has no consequence.

What to carry forward

Make causal explanations testable, compare the right outcomes and retain enough history for another person to understand and challenge the conclusion.