You are correct, however If I knew some shells have little or no powder, I would weigh ten that I knew were fine, with all other things as equal as you can expect. Then weigh the flat that is known to have an issue and would take out any that was off by 10 or 12 grains.
It's not a perfect system, but you'd limit the odds of shooting shells with no powder.
That being said, my more likely solution is curse at myself for ten minutes, get all pissed off, put the shells I messed up over in the corner and forget about them.
If the problem is truly no powder (since the original problem was powder clumping this might not be true) then your approach is easily implemented. I used Recoil Sissy's numbers from above to run a quick analysis. With a mean of 684.64 and standard deviation of 3.432 (from the 10 shells listed - it would be nice to have a little larger sample but 10 is OK for playing around) the chances of a shell that actually had the right amount of powder being 14 grains low are 23 in 1 million. Only 2 in a million will be 16 grains low.
At the same time, the false negative rate (keeping a bad shell) can be calculated by looking at the chances a shell with no powder would have so much excess weight in the other components that it would seem to weigh the right amount. Using the same numbers, the chance a no-powder shell would come in at the mean is again practically zero (1 in 10 million). What is more important is choosing a threshold for tossing suspect shells. Using the data we are playing with, I tried to choose a point that made good sense (again assuming there is no powder in some shells and 18 grains is the proper amount of powder). Depending on personal preferences, here are some examples.
1. Toss any shell that is 10 grains or more below the mean. This would give a false positive rate of 0.2% (this is the chance of throwing away a good shell). At the same time, the false negative rate is 1% (chance of keeping a bad shell).
2. Toss any shell that is 8 grains or more below the mean. This would give a false positive rate of 1% (this is the chance of throwing away a good shell). At the same time, the false negative rate is 0.2% (chance of keeping a bad shell).
3. Toss any shell that is 6 grains or more below the mean. This would give a false positive rate of 4% (this is the chance of throwing away a good shell). At the same time, the false negative rate is 235 in 1 million (chance of keeping a bad shell).
Since the consequence of a false negative is a "poof" a chance of keeping a bad shell isn't too distasteful. At the same time, tossing a good shell or two won't hurt too badly either. I'd probably rather toss more good ones than sit and wonder when I'll embarrass myself. If the problem was overloads, I'd tend to toss a lot more to avoid any chance of bursting a barrel.
Of course, the numbers would change for the OP because the variation could be different for his reloader/components. The real problem would occur if instead of powder/no powder there is something more complicated going on (which, with powder clumping, is certainly possible). If some shells have no powder, some have short throws, and some are overloaded this is a much more complicated problem and I'd just toss the bunch rather than spend the time and effort trying to set limits. This approach would work well for those of us who forget to fill the powder bottle sometimes...
I would be sure to weigh a bunch of empties and stike an average as the gentleman did above. I would be sure to use all one brand of shell.
I'll bet no conclusive facts will prove anything.