I don't think Peter Klausler did exactly what I'm suggesting. I should be more specific.

Say that you had a little program to type sample of text into. The program would keep track of the time taken in between each key stroke. You could ask a bunch of people to type the text sample as quick as they could, and record the key delays. If you knew which keyboard layout the subjects used, you could map the pressed characters to the location of the pressed keys. That way, the data would be more relevant and much less dependent on the keyboard layout used. The data would vary for subjects with different muscle memory, finger size, etc. However, on average, I feel as though it would be meaningful.

Hmm... it would also be good to normalize each user's timings. In other words, divide all of each user's timings by how long they took to type the entire text.

There would be many key combos in which the hand movement is the same on two different layouts, but the associated letters are different. That means that the combo will, on average, have been practiced more or less for each layout due to the frequency of that pair of letters. If the typists are well practiced, and averages are taken, this difference would be less. In fact, if many different keyboard layouts were used, the averages would be even better.

To resolve the specific case I mentioned in my initial post we would extract all timings associated with two keys pressed adjacent on the keyboard, next to each other in the same word, and pressed by the same hand. Perhaps there could be more restrictions as well. Then take the average. Then extract all timings associated with two keys pressed, one by each hand, pressed one immediately after the other, in the same word. If it's true that adjacently pressed keys happens faster than hand-alternatingly pressed keys, then the data may confirm it.

Klausler says, "I constructed a complicated function that measures the amount of "work" needed to touch-type a given text with a given layout." He goes on to explain his made up function. If he instead recorded key timings and based his function on how fast people actually type, that would be really cool!

Awesome!  : )

That's pretty cool! It does seem like a much lower keypress-time filter would do. And also, it seems like he could have analyzed the data from more perspectives, with more data, like you said. I like his idea... perhaps if I'm very bored one day, I'll modify it to keep zxcv where they are and give better ratings to qwerty-like layouts...