I just go by what Cherry datasheets tell. I don't have any interest in guesswork or measuring myself. I would assume their information about their own product should be correct.
Statistics isn't guesswork. Their data sheets are likely accurate but probably only report the mean force. What's interesting to me is the standard deviation. Since if their tolerances are high, it could lead to inconsistent typing experiences across different boards with the same kind of switch.
Their data sheets have 2 measures, actuation and bottom with a force graph. I can't find sheet for a couple switch though, and have some for ones that aren't made anymore, like clicky greys... which I have never even seen. 
Cherry reports the tolerance on switch for ±20cN so it is pretty loose really.
Standard Deviation: You would need a sample size of at least 50 keyboards with the same switch type to calculate a standard deviation with a high level of confidence. On the other hand, you can reverse calculate the standard deviation for a pretty good "guess-timation", and save the money from buying 50 keyboards, and the time it will take to test every switch x 50 keyboards. If you take IvanIvanovich's 20 cN maximum differential, then you can work back under the assumption that the cherry switch manufacturing process conforms to a normal distribution (which it should, since there is no reason to believe it would be skewed.) i.e., +20cN will be your high 3 sigma, and -20cN will be your low 3 sigma.
I don't know where you got that 50 number. If you're thinking of the made up cutoff at which we tell students they can use a normal distribution instead of a t when computing confidence intervals and hypothesis tests, that number is 30. It's also not entirely relevant here. If you wanted to compute a confidence interval for the variance or test the hypothesis that it is ~20/3 you'd be using a chi-squared distribution anyway(still assuming normality).
While I agree that normality is reasonable assumption, why do you think 20cN is 3SDs? Is this something standard in manufacturing that I just don't know about?
Another big assumption people are making here is that each switch/spring is i.i.d (Independent and Identically Distributed). Under this model, for every spring produced we would just pick a number from a normal distribution and be done with it. However, I have my doubts as to this assumption. If I were to hazard a guess, there would be two main sources of variability. One would be intra batch, meaning that in every batch of springs they make there is some true mean actuation force that probably isn't exactly the overall mean and some small amount of variability. The other would be inter-batch. Since my guess is that springs are made in runs, you could could think of the mean actuation force for springs in a given batch was also sampled from a normal distribution. This would give you a nice hierarchical model (if you're a Bayesian).
Sorry if this seemed dashed off. I'm in a hurry. But it could be fun.
Well, i was trying to give you an easy way to do a guess-timation calculation, not an exact scientific number. So that we could assume for the sake of discussion that, in general 68% of all springs will have the stated mean actuation force +-5cN or something like that. And that most likely since +- 5cN is not noticeable by human fingers, so that you would only "feel" a difference in a very small number of switches (i.e., 2 or 3 sigma), and it would not affect the overall feel of a keyboard, which I think is what real-world experience tells us. Otherwise there would be threads on this forum about how funky Cherry keyboards feel.
As for a manufacturing standard on sigma, there is none, which is part of the problem. The max and min tolerances should be six sigma, but that is expensive, and the only company that uses such a tolerance, that I know of, is Carl Zeiss AG (the Swiss Camera lens company), and check out their prices. Otherwise, you are lucky to get 3 sigma, usually you get 2 sigma. As for the springs, I think they are all made in China, so who knows about the quality, or the tolerances. However, there should be a normal distribution among the springs, even between batches, and I am assuming Cherry has set some min and max tolerances and some confidence interval. Of course this assumes no cheating by the manufacturer, or by his suppliers, and no defective input materiel (which is a dubious real-world assumption, especially in China, I know). As for the sample size of 50, that is what I have seen in real world testing. For some reason nobody uses the minimum of 30.
