Reposted from my cache server. I like my squid :)
----------------------------------------------------------
I finally had a chance to examine my old timing data, and I believe I can now answer my own original question, including a couple of pretty significant surprises.
Ok, so first off, I goofed a bit - my old timing data was from when I was using an ergo (non-kinesis) keyboard! Meaning, the MKJ rotation and the G/V swap is ideal as far as speed is concerned FOR NON-KINESIS keyboards. What's really mind blowing is that mirrored-colemak should be more comfortable for me than normal colemak on a normal keyboard due to better placements of J, V, M and B. (This is going strictly from my timing data and how much finger stretch when row jumping is involved.)
However, using a grid-type keyboard, Shai's layout modified with a G/V swap is fastest for me, even with the increased row jumping. No rotation necessary! Isn't that crazy? Mirrored colemak is (almost) ready-made for grid keyboards. :)
----------------------------------------
SpeedMorph, regarding your scoring system - I do see a few problems with it, but it's subjective. The cost system you've designed seems a bit arbitrary - maybe have it based on more concrete data than just distance travelled. You should also tune your program to detect awkward finger positions between common phrases, not just between letter pairs (for instance, I type "THERE ARE" or "IT IS" in one motion. My algorithm checks to see if there are row jumps among the words as well) but since you're keeping the 10 most common letters in home row, chances are that this additional check won't really be necessary. I also implemented a subroutine to optimise for alternate fingering as well based on my typing habits. You could try that to see if you shave off a few seconds there.
I also think it's good to have a clear goal in mind. Are you looking to maximise speed? Or are you more concerned about comfort? If you're using a corpus that includes C code, and writing style that is dissimilar to your own, then it's going to skew your results. It seems like your corpus is targeted for the all around coder/typist with comfort in mind - so sacrifices in any one area will be made while you run your evaluation program.
Oh yeah, letter frequency is not the same as letter distribution. Just because your letter frequency matches with an online source doesn't mean the corpus is good. My advice - ditch the C code, and add more sources that exhibit similar writing style to your own and focus on a specific goal.
Anyway, here are the final layouts based on my data and suggested by my program to be optimised versions of colemak (***remember, the layouts below are MIRRORED*** - as if you're looking up through the keyboard from below):
FOR "MS ERGO" AND STAGGERED KEYBOARDS:
Q W F P K M L U Y '
A R S T D H N E I O
Z X C G B J V ; , .
FOR "GRID" KEYBOARDS (swap V and K for slightly lower "same-finger"):
Q W F P V J L U Y '
A R S T D H N E I O
Z X C G B K M ; , .Well, that's it for now. Time to start ramping up my speed. Currently stuck @ 65wpm.
EDIT: W00T! Just hit 70WPM! Close to tying my old colemak record, and it only took me 6 weeks to hit 70 this time! My alternate fingering is working! :)