I agree with statdoug. You should take a minute to prepare a
total variance equation to gain understanding of total variation you are measuring, and how to
eliminate or reduce to an insignificant level those factors you do not wish to measure. Since in this case you are judging the
measurement, you need to eliminate the
process variance - which, as statdoug suggested, can be done by using the same specimens for the evaluation. But are there more variations that need to be eliminated to focus on the issue? Only you can tell.
As far as distributions, you are currently measuring the sum of all distributions participating in the total variance equation - which will be multimodal, but most software will attempt to dump it into normal. Typically, once you get all other distributions eliminate, the measurement system variation should be a normal distribution, as that sources of variation are a natural source. The mix of the distributions from the other sources of variation will make it wider and 'lumpier' than it truly is.
Another suggestion is to supply sample charts for the machines in question. There are a variety of charting methodologies for group charts, normalizing data (as in short run type charts), etc. Too many to describe - but we can try some and let you consider them.
