A Gage R&R asks one question of a measurement: how much of the spread in the numbers is the thing being measured, and how much is the gauge and the person holding it? Factories ask it before a new caliper or test rig is trusted to pass or reject parts. The new Gage R&R tool answers it by the ANOVA method, for crossed studies (every operator measures every part) and nested ones (destructive tests, where each operator gets their own parts). It runs in the page, the command line or an agent call, and nothing is uploaded.
A stopwatch is a gauge
The tool needed a real study for its example, not an invented caliper. A benchmark harness turns a function into a number, which makes it a gauge, and this one could be measured on the spot. So: ten "parts" (sorting copies of arrays of 400 to 4,000 random numbers), three "operators" (timing harnesses), three trials each, 90 readings in shuffled order (eval/benchmark-study.mjs):
single: one call between twoperformance.now()readings;loop50: fifty calls, total divided by fifty;median15: the median of fifteen single calls timed withprocess.hrtime.bigint().
With the part × operator interaction kept in the model:
| Source | %study variation |
|---|---|
| Gauge R&R | 22.5 |
| Repeatability | 15.7 |
| Reproducibility | 16.1 |
| Part-to-part | 97.4 |
That gives 6 distinct categories. By the AIAG manual's usual bands, that is "may be acceptable": between 10% and 30%.
The reproducibility is nearly all one harness. single averages 101.5 µs a call; loop50 and median15 read 83.4 and 85.0. Timing a single short call adds the timer's own cost to every reading, so the harness reads high on every part. The interaction plot shows it as a line sitting above the other two. Drop single and gauge R&R falls to 14.7%, with 9 distinct categories.
These are one container's timings on one day. A rerun gives different numbers, which is the point of a gauge study. The file the page ships is the one measured for it.
The number nobody prints: how sure the study is
Gauge reports usually print 22.5% as if it were a fact. But it comes from ten parts and three operators, and with three operators reproducibility rests on two degrees of freedom. The tool gives every variance component a confidence interval: an exact chi-square interval for repeatability, and modified large-sample intervals (Graybill & Wang 1980; Ting et al. 1990, reviewed for gauge studies by Burdick, Borror & Montgomery 2003) for the rest. For the benchmark, the gauge R&R standard deviation is 14.7 µs, with a 95% interval of 11.6 to 64.1 µs. The study can say that the gauge is not perfect. It cannot say whether it is merely marginal or badly off.
Formulas like these are easy to get subtly wrong, so the tool checks the intervals by simulation instead of trusting them. It drew 2,000 studies from the random-effects model, with known variances, for each of seven designs: crossed and nested, 5 to 20 parts, some with no operator or interaction effect at all. Each interval was then checked for whether it contained the true value. Every component's 95% interval did so 94.5% to 97.7% of the time (eval/coverage-output.txt). That is slightly conservative, which is how MLS intervals are known to behave.
One choice goes beyond the papers. The total gauge interval's lower end is not allowed to fall below repeatability's, because gauge variance cannot be smaller than one of its own parts. The nested example needed it: there the unbounded MLS lower limit was 0, while repeatability's was 1.56. The change can only raise coverage, and the simulation agrees.
Checked against
- Sums of squares from statsmodels 0.15.0 OLS ANOVA on eight generated studies, variance components from numpy solving the expected mean squares, and p-values and quantiles from scipy 1.18.1, all to 1e-9 (
reference.test.mjs). - A 2 × 2 × 2 study worked by hand. The first draft of that test was wrong: the hand arithmetic, not the engine. That is a fair reminder of why the Python reference exists.
What it does not do yet
It handles balanced studies only, and treats operators as random. It does not yet do the average-and-range method, attribute agreement, or bias and linearity studies. The next useful addition is probably a bias and linearity study, the other half of a typical measurement system analysis package.
Not affiliated with the cited authors, the AIAG or any statistics vendor. Statistical estimates, not engineering certification. Written with AI assistance.