CQT common mistakes
CQT · Understand the traps and how to avoid them.
Answers a time-order question with the wrong basic tool - for example, asked whether a defect rate has trended upward over the last twelve weeks, builds a histogram of weekly counts instead of plotting the weeks in sequence; or uses a Pareto chart for continuous measurement data.
The 'seven basic tools' are memorized as a list, so candidates pattern-match on the word 'data' instead of on the question being asked, and the histogram feels like the safe default chart.
What to do instead
Name the question type before picking the tool: distribution/shape → histogram or box plot; time order or trend → run chart or control chart; ranking of categories → Pareto; relationship between two variables → scatter diagram; process flow → flowchart; sources of variation → cause-and-effect. Then check the horizontal axis: a time/sequence axis means a run or control chart, never a histogram.
Misreads the Pareto diagram - takes the cumulative percentage from the left (count or cost) axis, or asserts '80% of the problems come from 20% of the causes' regardless of what the curve actually shows, or re-sorts categories into a second Pareto without keeping the same category definitions and 'Other' group.
The 80/20 phrase is learned as a slogan, dual-axis charts invite reading the wrong scale, and the catch-all 'Other' bar hides how the ranking was constructed.
What to do instead
Read bar heights on the left axis and cumulative points on the right (percent) axis, and state the breakpoint that is actually in the data rather than quoting 80/20. When comparing two Pareto charts (before/after), keep category definitions, order rules, and the 'Other' grouping identical; otherwise the comparison measures the regrouping, not the improvement.
Calculates capability with the wrong sigma - computes Cpk from the sample standard deviation of all individual readings, or mixes Cp/Cpk (within-subgroup sigma from R-bar/d2 or s-bar/c4) with Pp/Ppk (overall sample sigma), so the capability numbers contradict the control chart built from the same data.
Only one standard deviation formula is retained, and the within-subgroup (short-term) versus overall (long-term) distinction is skipped as a formula detail rather than understood as two different estimates of variation.
What to do instead
Describe the data structure first. With rational subgroups, Cp/Cpk use sigma-within = R-bar/d2 (or s-bar/c4); Pp/Ppk use the overall sample standard deviation. Use n-1 when data are a sample from a process, n only when the data are the whole population. Sanity checks: overall sigma should be ≥ within sigma, so Ppk ≤ Cpk, and the chart's control limits should be computed from the same subgroup statistics.
Confuses control limits with specification limits - draws the spec limits on an X-bar/R chart and calls a point 'out of control' because it exceeds spec, or calls a process 'in control' because every point sits inside the specification, and ignores run, trend, and zone rules entirely.
Both are horizontal lines on the same style of chart, and the idea that control limits are computed from the process's own variation (not imposed by a customer) is easy to recite but easy to drop under time pressure.
What to do instead
Say it explicitly: control limits are the voice of the process, computed from the data (X-bar ± A2·R-bar, R-bar·D3/D4, etc.); specification limits are the voice of the customer and are never drawn as limits on a variables chart. Then apply all signal rules, e.g., one point beyond 3 sigma, 2 of 3 beyond 2 sigma, 4 of 5 beyond 1 sigma, and 8 consecutive points on one side of the centerline.
Defines AQL as 'the maximum percent defective a lot may contain and still be accepted,' treats a rejected lot as inspector error, and swaps producer's risk (alpha, rejecting a good lot) with consumer's risk (beta, accepting a bad lot).
AQL sounds like a pass/fail cutoff or a quality target, while the operating-characteristic curve logic - a quality level with a specified probability of acceptance - is abstract and rarely practiced.
What to do instead
State AQL as the quality level at which lots have a high probability of acceptance (about 95%), so lots at AQL can still be rejected; that rejection is producer's risk (alpha). LTPD/RQL is the level with a low acceptance probability (about 10%), tied to consumer's risk (beta). When a question supplies risk percentages, sketch the OC curve or recall its shape before choosing, and never describe a sampled lot's quality as an exact percentage.
Treats a sample result as a precise measurement of lot quality ('2 defectives in a sample of 50, so the lot is 4% defective'), assumes 100% inspection guarantees zero defectives, and applies normal/tightened/reduced sampling tables without checking lot-size ranges, inspection level, or switching rules.
Acceptance sampling feels like measuring rather than like applying a decision rule, and the tables look mechanical, so the conditions surrounding them get skipped.
What to do instead
Phrase sample results as evidence for accepting or rejecting the lot, not as an estimate of lot quality (estimation needs a different sampling approach). Remember inspection is imperfect - visual inspection in particular has documented miss rates and fatigue effects - so 100% screening cannot promise zero defectives. Before using any plan, confirm lot size range, inspection level, AQL, and which of normal, tightened, or reduced applies.
Calls a gauge 'accurate' because its repeated readings agree with one another, blames 'poor repeatability' for variation between appraisers, and in a Gauge R&R confuses equipment variation (repeatability) with appraiser variation (reproducibility), or declares the gauge acceptable without reporting %GRR or ndc.
Accuracy/precision and repeatability/reproducibility are near-synonyms in everyday English, and the G&R arithmetic is memorized as formulas (EV from average range, AV from differences among appraiser averages) without the definitions behind them.
What to do instead
Anchor on definitions: accuracy is closeness to the reference value (bias); precision is closeness of repeated readings to each other. Repeatability = one appraiser, one gauge, same part, repeated trials (equipment variation); reproducibility = different appraisers using the same gauge (appraiser variation). Report %GRR and the number of distinct categories before judging, using the customary 10% / 30% guideline.
Uses 'calibration' interchangeably with 'adjustment' or 'verification,' treats an 'as-found out of tolerance' result as automatically requiring scrapping of the gauge, and decides pass/fail from the reading alone without considering the traceability chain and measurement uncertainty.
Everyday usage equates calibration with fixing an instrument, and the metrology definitions separate comparing an instrument against a standard from physically changing it - a distinction that disappears if only the procedure is memorized.
What to do instead
Keep three terms separate: calibration is comparison to a traceable standard with a recorded result; adjustment is a physical change to bring the instrument within tolerance; verification is checking that an item meets a stated requirement. Always record as-found before as-left, compare error plus expanded uncertainty against the tolerance, and if the gauge was out of tolerance, assess the effect on product measured since the last valid calibration before deciding disposition.
Offers rework, sorting, or 'retrain the operator' as the corrective action for a nonconformance, never separates containment from correction, and closes the issue without verifying that the action worked.
Immediately fixing the problem feels like solving it, and the standards' sequence (contain, review, determine cause, act on cause, verify effectiveness, update risks and documentation) gets compressed in memory into a single 'fix it' step.
What to do instead
Label each step explicitly: containment protects the customer; correction fixes this instance; root cause analysis (5 Whys, cause-and-effect, confirmed with data) identifies the cause; corrective action eliminates the cause; then verify effectiveness and update documents and risk assessments. If the answer changes product rather than the system that produced it, it is a correction, not corrective action.
Treats verification and validation as synonyms, and confuses a procedure or work instruction with a record - for example, answering that a completed inspection sheet is controlled as a procedure, or that a work instruction is retained as evidence that the activity was performed.
Both words mean 'checking' in ordinary language, and the document hierarchy (manual, procedure, work instruction, form, record) blurs because all of it looks like paperwork.
What to do instead
Verification is objective evidence that specified requirements were met (did we build it right - inspection, dimensional checks, calibration); validation is evidence that the intended use is satisfied (did we build the right thing - process validation, design validation). Documents tell you what to do and are controlled by approval and revision; records are evidence of what was done and are controlled by retention, legibility, and protection.