EPKL has Shai's latest (D⇔P) mod as well as SteveP's D>P>G mod in the Remap file, described as a "poor man's DH" variant leaving the bottom left half row alone. To use one of those you may rename which variant bears the name "Colemak-Curl-DH" in the Remap file. The help images won't automatically update but it's possible to make new ones.

Q W F D G   J L U Y ;
A R S T P   M N E I O    Shai's 1st mod; Cmk-DpH [QU-SC swap not shown here]
Z X C V B   K H , . /

As for the other variant, I haven't put it in there. It plays with some of the same ideas as Nyfee's mods, such as P⇔M and V/B>K. I've yet to see a more thorough analysis of its merits.

Q W F M V   J L U Y ;
A R S T G   P N E I O    Shai's 2nd mod
Z X C D K   B H , . /

These mods as analyzed by the Colemakmods analyzer in a 3×10 core matrix as above in this post, didn't fare so well:

  Cmk-DH+swaps/loops: bas.ef. sfb-ef. nfb-ef. tot.ef. SFB%      Δbas.e  Δsfb-e  Δtot.e
  =========================================================    ========================
  Cmk-DH            : 1.630   0.047   0.010   1.687   1.521      0       0       0
  ColemaQ           : 1.645   0.031   0.011   1.687   0.974     +0.015  -0.016   0.000 (VK      + /, + Q;)
  Colemak-QI   noQU : 1.624   0.040   0.017   1.681   1.275     -0.006  -0.007  -0.006 (PM + BK + WFL)
  Shai's Cmk-DpH    : 1.657   0.047   0.010   1.714   1.521     +0.027   0.000  +0.027 (DP instead of DvBG)
  Shai's 2nd mod    : 1.622   0.054   0.010   1.686   1.671     -0.008  +0.007  -0.001 (PM + VBK)
  Cmk-DH +VK only   : 1.630   0.041   0.009   1.679   1.311      0.000  -0.006  -0.008 !!!
  ISRT/IndyRad noQU : 1.625   0.024   0.010   1.659   0.777     -0.005  -0.023  -0.028 !!!

Q W F P B   J L U Y ;
A R S T G   M N E I O    Cmk-DH (3×10)
Z X C D V   K H , . /

It's a fairly simple analyzer, so maybe someone has more relevant analysis!

Yes, I know. For the purpose of analysis, however, I didn't include the QU-SC swap since such a change would lead to really unfair comparisons the way I set it up. It may technically be considered a Sym-Qu mod anyway, separate from the rest of the layout. You could add the same swap to QWERTY or vanilla Colemak if you wanted to, just like it is with the rest of the Sym mod. Mod modularity, I love it.  ( のvの) c[_]

I should've shown that in the diagram text, however. Sorry about that oversight.

This chart is not final either ... the relative positions jumped a bit in the last trial run. The dot for Shai is under the red one for Colemak.
(grabbed wrong image)

The discussion is about Colemak vs Shai, so I only included those two from the spreadsheet. There are 362 layouts currently analyzed, but I need to remove a few dups. There's a lot of Colemak flavours.

Colemak is one of the "defaults" on the chart, normally a different "layout of interest"  would be given the red dot. I need to modify the page to filter out the legends for the "defaults" if one of them is the "layout of interest". I meant to grab the chart for Shai, in which case it would have been the red not, with Colemak again invisible under it.

If I make the dots smaller then they become diamond-shaped. Currently drawn with PHP GD library, should maybe switch to making it an .svg image instead.

Job is currently rerunning, will post update later. It takes several hours.

Cheers, Ian

Variants and scores.

There's also madness in your method!  ( ͡~  ͜ʖ ͡°) Thanks for the update.

Out of interest: How does your Colemak-mod-Ian-2 look?

I think it's this, looks messy but maybe there's a clear explanation in it:
https://www.keyboard-design.com/index.html

I'd love to see your comments on these analyses, Steve and Shai!

OK, yes that rings a bell, I have looked at iandoug's pages before. So basically it's based on the patorjk analyzer, which itself is based largely on finger movement distance, but there are also some changes by "Den".

Looking at "Best layouts" section, I notice it says something about taking account of vertical vs horizontal movement, which is something I've done in my fork too, so that's good. Judging by this page though, the "Colemak Mod-DH" layout isn't right as the angle mod is missing!  I don't know whether this analyzer distinguishes between the Qwerty N/M positions for example, as both are equally distant from the respective home key (J), but obviously in reality the M spot is better than the N.

I also don't understand this section:
For a given input text, we assume that a finger would need to move one key width (plus down and up and back) for each character in the text. So the distance is calculated as (number of characters * (19 + 4 + 4 + 19) mm). This is then compared to the actual distance measured, and turned into a percentage.

I'd question the usefulness of the word-based metrics, as it's just a count of how many words you can type with a specific subset of keys. It's an interesting side effect to note, but not really that useful as a comparison tool IMO, which is acknowledged to be fair.

Would be nice if the algorithm being used was defined more clearly.

Picture worth 1000.

Yes, that's Colemak-DH with the Angle-Z mod which is the most common one for ANSI boards. Cmk-DH-ANSI, if you wish.

Please ignore spreadsheets and graphs posted so far... current results rather different.

Previous Finger-based metrics (ie the scores from KLATest) were based on a short 5k English and 10k code input. I did this just to filter out dups and dev versions etc.

I have now rerun the tests with proper 1 MB input for each, Colemak score looking better. Even better than mod-DH ...

I need to double-check everything before posting results.

I also changed how I did the word metrics (twice!). Was initially just counting "words" but clearly "verylongword" should count for more than "short". So I tried counting letters instead. However, since the word source is the Linux word list (from a few years ago), there are many words that have things like -nesses at the end which bump up the length but use the same letters. So I switched to counting unique letters per word instead.

Cheers, Ian

Personally I don't think the V-K swap is anywhere close to being worthwhile. V and K are both rare so we are talking about very minor benefits, if any. It avoids same-finger LK, but at a cost of doubling the number of keys that swap hands (from 2 to 4), and it moves V which (assuming angle mod), is unchanged from Qwerty. And it also splits up XCV even more, so that paste is now on the other hand. Even if analyzers rate it as slightly better, the non-algorithmic drawbacks cause me to be highly skeptical.