Skip to content

Gauge Repeatability and Reproducibility Study

A measurement audit — instantiates Tolerance Stack Management

Separates the variation that comes from the parts from the variation that comes from measuring them, so that a stack analysis is not silently built on the noise of its own gauges.

Every tolerance band and every stack computation rests on measured numbers, and a measurement is itself a small stack of deviations — the gauge's own repeatability plus the disagreement between operators. Gauge Repeatability and Reproducibility Study quantifies exactly that: it partitions the total observed variation into the part it should reflect (true part-to-part difference) and the part it should not (the measurement system's own scatter). Its defining role among the siblings is that it audits the instrument, not the design — it asks whether the numbers feeding the whole stack are trustworthy in the first place, and it can reveal that measurement noise is itself one of the largest apparent contributors to a fit problem. It is the check that keeps a tolerance analysis from being precise about garbage.

Example

A shop is chasing intermittent assembly failures and suspects a machined shaft is drifting out of its ±0.01 mm diameter band. Before re-cutting the process, they run a Gauge R&R on the measurement itself: ten shafts, three operators, each measuring each shaft twice with the shop calipers, in randomized order. Analysis of the spread shows that repeatability (the same operator remeasuring) is fine, but reproducibility (operator-to-operator) is large — one inspector seats the caliper on a burr the others clear. The measurement system alone accounts for roughly 35% of the total observed variation, well past the ~30% rule-of-thumb ceiling for a marginal gauge, and it resolves only a handful of distinct part categories. The finding flips the investigation: the shafts were largely in tolerance; the "drift" was mostly measurement disagreement. Fixing the measurement method removes an apparent contributor that no design change would ever have addressed.

How it works

The study is a small designed experiment: several parts, several operators, repeated measurements, deliberately randomized so bias can't hide in order. The spread is decomposed into repeatability (equipment scatter on a repeat) and reproducibility (systematic difference between operators), combined into the measurement system's total variation, and expressed as a fraction of either the tolerance band or the total observed variation. What distinguishes it from every calculating sibling is that it treats the measuring as the object of study — it is variation analysis pointed at the gauge rather than at the design.

Tuning parameters

  • Parts / operators / trials — the size of the crossed design. More of each tightens the estimate but costs time; too few and the study can't tell equipment noise from operator disagreement.
  • Part selection span — whether the sample parts cover the full process range or cluster near nominal. Parts that span the range separate gauge noise from real spread; clustered parts inflate the gauge's apparent share.
  • Acceptance basis — whether %GRR is judged against the tolerance (fitness to gauge this spec) or against total process variation (fitness to study this process). They answer different questions and can disagree.
  • Randomization and blinding — how hard the study hides part identity and order from operators; weak blinding lets expectation leak into the numbers.
  • Resolution check — the number of distinct categories the gauge can actually separate; too few means the instrument can't see the variation the stack cares about.

When it helps, and when it misleads

Its strength is that it protects every downstream number: it distinguishes a real fit problem from an artifact of the gauge, and it can unmask measurement as a dominant, previously-invisible contributor — the classic case where the "process is drifting" turns out to be the ruler moving. It is the measurement-systems-analysis discipline that underwrites trusting a tolerance inventory at all.[1]

It misleads when its result is over-read: passing a Gauge R&R proves the gauge can discriminate, not that it is accurate — a repeatable, reproducible instrument can still be biased against a true standard, which calibration, not R&R, catches. It also assumes the study parts represent the real population; a sample bunched near nominal will flatter the gauge. And a study run once and filed ignores drift as the instrument wears. The discipline is to pair R&R with calibration to a traceable standard, sample across the true range, and re-run periodically rather than treating one green result as permanent.

How it implements the components

  • measurement_system_check — it is the measurement-system check: it validates that the numbers feeding the stack reflect the parts rather than the act of measuring them.
  • critical_contributor_map — by sizing measurement scatter against the other contributors, it reveals whether the gauge itself is one of the few dominant sources of apparent variation, redirecting effort accordingly.

It does not model how design deviations accumulate, nor set or check the fit limit — the accumulation is done by Statistical Tolerance Analysis and its calculating siblings, and the fit check by Integration Acceptance Test.

Notes

Gauge R&R measures precision, not accuracy: it tells you your measurements agree with each other, not that they agree with truth. Bias against a traceable standard is a separate check, and confusing the two is the most common way a green R&R gives false comfort to a stack that is quietly measured off-center.

References

[1] Measurement Systems Analysis treats the measurement process as a source of variation to be quantified and controlled; Gauge R&R is its workhorse study, reporting the measurement system's share of total variation (commonly the %GRR statistic) and the number of distinct categories the gauge can resolve. ↩