A cap gauge can be calibrated, in tolerance and still produce disagreement. Two operators check the same batch of E27 caps on the same morning and reject different parts. The gauge is not necessarily at fault and the operators are not necessarily wrong: the measurement system as a whole is not capable, and the only way to find that out before a customer does is a study that separates the gauge from the person holding it.
The classical gauge R&R study assumes the instrument returns a number. A cap gauge returns a verdict — enters, does not enter — and the information in that verdict is one bit. Running a variable study on it produces a repeatability figure that describes how consistently an operator pushes, not how well the gauge measures. What the gauge is actually doing is comparing a cap dimension against a limit, and the limit is fixed by a 7006 sheet under IEC 60061-3.
That does not make the study useless. It changes the design. The variable version tells you whether the gauge can discriminate between good and marginal caps; the attribute version tells you whether the pass/fail decision is stable across people and repeats. A gauge room needs both, in that order.
| Element | Value | Note |
| Parts | 10 caps spanning the tolerance | Include two at each limit, not ten good ones |
| Appraisers | 3 | People who normally do the job |
| Trials | 3 per appraiser | 90 readings in total, parts randomised |
| Acceptable | %GRR below 10 percent | Gauge error small against tolerance |
| Marginal | 10 to 30 percent | Acceptable only with a documented reason |
| Unacceptable | Above 30 percent | The decision is partly noise |
| Room | 20 degrees C, soaked 2 hours | Steel and brass expand at different rates |
For a go/no-go gauge the defensible design is 30 caps, of which roughly half are conforming and half are deliberately at or just outside the limit, each checked three times by three appraisers. That is 270 verdicts and it produces three numbers worth reporting: agreement within an appraiser, agreement between appraisers, and agreement with the reference decision established by measurement.
A common acceptance line is 0.75 for the kappa statistic, which is the agreement corrected for chance. Below that, the verdict depends on who is holding the gauge, and no amount of gauge calibration fixes it. The usual causes, in order of frequency, are a worn stop face that lets each operator push to a different depth, a cap that rocks because it is supported on a thread rather than on the shell, and a procedure that does not say how hard to push.
Bias and repeatability fail differently. High bias with low repeatability means everyone is consistent and everyone is wrong, which points at the gauge or the reference, and is the easier case: re-verify against the 7006 sheet and the setting master. Low bias with high repeatability across appraisers means the gauge is fine and the technique is not, which is a training and fixturing problem. Both high at once usually means the caps themselves are inconsistent, and the study has found a moulding problem rather than a measurement problem.
Run the study on a gauge that has just been verified, so the result describes the process rather than an overdue instrument. The E27 go and no-go gauge pair is a useful subject because both members are involved and they fail in opposite directions. For the dimensional side, reading a 7006 gauge sheet sets out which dimension each member is checking.