9.4 Inspection Errors & Human Factors

Why Inspection Errors Are a Blueprint Topic

Every measurement discussed so far assumed the inspector executed it correctly. Body of Knowledge topic III.C.2 exists because that assumption fails often enough to matter. An inspection error is any outcome in which the recorded result does not reflect the true condition of the part, and the error can come from the instrument, the setup, the environment, or — most often — the person.

The Two Error Types

The same two error types you met in acceptance sampling apply to the individual inspector. Type II errors are the more serious because a false reject is eventually caught downstream — someone re-measures the "bad" part and argues — while a false accept simply ships.

Attribute agreement analysis is the study that quantifies both. Multiple appraisers evaluate the same set of parts, including known-good and known-bad units, several times each. The analysis reports each appraiser's agreement with the standard, agreement between appraisers, and repeatability with themselves, often summarized as Cohen's or Fleiss's kappa. A kappa above roughly 0.75 is generally considered good agreement, and below about 0.40 is poor.

---

Bias

Bias in the human-factors sense is any systematic tendency to record results in a particular direction. It is distinct from instrument bias (the metrological offset covered in measurement system analysis), though the two produce identical-looking data.

Common forms:

• Expectation bias. The inspector knows what the dimension "should" be and reads what is expected. A vernier that could plausibly be read as 1.247 or 1.248 gets read as the one matching the print.
• Anchoring. The first part of a run measured 0.7503, so every subsequent reading gravitates toward 0.7503.
• Source bias. Material from a supplier with a good reputation, or from an internal department with organizational clout, is examined less critically than identical material from a troubled source.
• Outcome bias. Knowing that a rejection will stop the line, blow the ship date, or trigger an argument with a supervisor tilts marginal calls toward acceptance.

Countermeasures are structural, not motivational. Telling an inspector to "be objective" does nothing. What works: blind inspection, in which the appraiser does not know the expected value, the supplier, or the prior result; removing specification limits from the raw data sheet so the inspector records a number rather than a verdict; automated data capture direct from the gage so the reading cannot be edited on its way to the record; and independent re-inspection of a sample of accepted parts by a second appraiser.

---

Flinching

Flinching is the single most exam-relevant inspection error because it has a precise definition that candidates routinely blur into generic bias.

Flinching is recording a borderline measurement as being inside the specification limit when the actual reading is at or just outside it.

The inspector measures 0.5011 in against a 0.4990–0.5010 in requirement, hesitates, and writes 0.5010. Or reads 0.4988 and records 0.4990. The reading is not fabricated wholesale; it is nudged just far enough to avoid the consequences of a rejection. Flinching is sometimes called "rounding to the limit" or "gundecking" a marginal reading.

Its data signature is unmistakable and is exactly what an auditor looks for: a histogram of recorded values with an unnatural spike exactly at the specification limit and a suspicious absence of values just beyond it. A truly capable process produces a smooth distribution; a flinched data set produces a cliff at the limit. The same cliff shape appears when 100% sorting has removed out-of-tolerance parts, so the auditor's next question is whether sorting was performed or whether the numbers were shaved.

Countermeasures are the same structural ones used against bias, plus two specific to flinching: do not print the tolerance on the data collection sheet, and capture the reading electronically from the gage so there is no human transcription step in which a nudge can occur.

---

Fatigue

Visual and repetitive inspection performance degrades measurably with time on task. Research and long industrial experience converge on a hard number that the CQI exam expects you to know: sustained 100% manual visual inspection is typically only about 80% effective. Roughly one defect in five escapes even a diligent inspector, and 100% inspection performed twice does not reach 100% because the same defects tend to be missed both times.

Contributors and countermeasures:

The strategic lesson is the one Deming made famous: because screening is only about 80% effective, the answer to escaping defects is process control and error-proofing, not more inspection.

---

Distraction

Interruption during a measurement sequence causes skipped steps, lost place in a sequence of characteristics, and mis-transcribed values. A CMM operator interrupted mid-alignment may resume with an incomplete datum setup. An inspector interrupted while filling in a 40-characteristic first article report may skip a line and shift every subsequent entry.

Countermeasures: designate a low-traffic, low-noise inspection area; adopt a restart rule requiring the inspector to return to the last completed and recorded characteristic rather than resuming from memory; use check-off sequences so the record itself shows where the work stopped; and control phone and radio use at the bench.

---

Poor Time Management

Time pressure changes inspection behavior in predictable ways. Work batched to the end of a shift gets measured under pressure. Sampling that should be spread across a run gets performed all at once on the last parts made, destroying the sample's representativeness. A patrol inspection scheduled hourly gets performed twice at the end of the shift and recorded as though it happened on schedule — which is a falsified record, not a scheduling problem.

Countermeasures: build inspection time into the routing and the standard work; schedule patrol inspections at defined intervals with electronic timestamps; and set clear escalation rules so an inspector who cannot complete required checks reports it rather than compressing them.

---

Setup Errors

Setup errors produce results that are wrong by a consistent amount across an entire lot, which makes them costly.

• Wrong master or wrong standard. Zeroing a comparator on the wrong gage block stack shifts every reading by the difference.
• Failure to zero. Not checking a micrometer's zero on its standard before use.
• Wrong datum or wrong setup orientation. Measuring from the wrong face makes every derived location wrong.
• Contamination. A chip, burr, or film of coolant on an anvil, a gage block wringing surface, or a surface plate.
• Inadequate thermal soak. Measuring before the part and gage have reached the 20 °C (68 °F) standard reference temperature.
• Wrong scale or unit. Reading the inch scale on a dual-scale instrument when the print is metric, or reading the wrong vernier line.
• Parallax. Viewing a dial or scale from an angle rather than square to it.
• Excessive or inconsistent measuring force. Overriding a micrometer ratchet or thimble friction stop.
• Incorrect part support. A long shaft supported at the wrong points sags; supporting at the Airy points minimizes end-slope error for a straightness measurement and the Bessel points minimize length change.

The universal countermeasure is a documented setup procedure with a verification step: measure a known master or check standard after completing the setup and before measuring production, and record that verification.

---

Cosine Error

Cosine error arises whenever the measuring instrument's sensitive direction is not aligned with the direction of the actual displacement. For a lever-type dial test indicator inclined at angle theta to the direction of workpiece movement, only the component along the stylus's sensitive direction is registered:

$T = \frac{M}{\cos(\theta)}$

where M is the dial reading and T is the true workpiece movement. Because the cosine of any nonzero angle is less than one, a tilted lever-type indicator under-reads the true displacement. Section 3.1 develops this geometry in full; what matters here is the practical rule and the magnitude:

The working rule is to keep the stylus within about 15 degrees of the workpiece surface, which holds the error under about 3.5% and is usually inside the measurement uncertainty budget. Beyond that, either reposition the indicator using its swivel or apply the cosine correction and document it.

Cosine error is not confined to test indicators. It appears whenever a CMM probes a surface away from the surface normal, whenever a laser interferometer's beam is not parallel to the axis of travel, and whenever a scale or straightedge is not aligned with the measurement direction.

---

Real Shop Scenario

A machining cell has run a 0.3750 in ± 0.0005 in journal diameter for six months at an apparent process capability index of 1.9. A customer returns three shafts measuring 0.3757 in. The quality engineer pulls two years of inspection records and builds a histogram of the recorded values.

The histogram is smooth and bell-shaped from 0.3745 to 0.3754, then shows a sharp spike of 31 readings at exactly 0.3755 — the upper limit — and not one recorded value above it. The process standard deviation implied by the smooth part of the curve says roughly 20 readings should have exceeded 0.3755 over that period.

There is no sorting operation on this part, so parts above the limit were never physically removed. The cliff is a flinching signature: marginal readings were being recorded at the limit rather than beyond it. The corrective action was structural — the data sheet was reissued without the tolerance printed on it, the bench comparator was connected to direct electronic data capture, and an attribute agreement analysis was run across the four inspectors on the cell.

---

Common Exam Traps

• Flinching has a specific meaning. It is recording a borderline value as in-tolerance, not general carelessness and not deliberate fraud on a grand scale.
• Type II is the dangerous error. False accept ships; false reject is usually caught internally.
• 100% inspection is not 100% effective. Roughly 80% is the working figure, and inspecting twice does not fix it.
• A tilted lever indicator under-reads. Divide by the cosine to get the true value; do not multiply.
• A spike at the specification limit has two possible causes. Flinching, or legitimate 100% sorting that removed the out-of-tolerance tail. Determine which before writing the finding.
• Countermeasures must be structural. Retraining and exhortation do not remove bias; removing the information that drives the bias does.