The evolution of P/F vs T/K raises important questions

[Jorge_Stolfi]

...

Sure, for each single one of my observations there are other possible explanations than the claim "p/f are fancy versions of k/t and/or ke/te".  But, taken together, they don't seem to leave any other option.

Latin alphabet 'p' 'q' 'b' and 'd' also look graphically to be a single class of glyphs (one circle plus one vertical bar), even more than EVA k/t/p/f do. Would you deduce for instance that Latin 'p' 'q' can be replacements for 'b' and 'd'?

That would be stupid, but I would take the similarity between "C" and "c", "F" and "f", "I" and "i", "J" and "j", "K" and "k", "M" and "m", "O" and "o", "P" and "p", "S" and "s", "U" and "u", "V" and "v", "X" and "x", "Y" and "y", and "Z" and "z" as hints that all capital letters may be just fancy versions of the lowercase ones. 

1. The first glyph of a parag is very often a puff.
   
2. In some sections, 4-6% of the glyphs on every headline, besides the initial, are puffs.
   
3. Apart from the head lines, puffs are rare -- maybe 2% of the words have them.
   
4. They are rare even on the second lines of parags that have very short head lines.
   

This only establishes that p/f are common only in a limited section of the manuscript, noone doubts that, but it does not follow that they are replacements for k/t. You should at least show that where p/f are common k/t are rare, this would be a more promising evidence.

Methinks that it is you who should show that k/t are just as common on head lines as they are in the text.

But, given "In every word structure model, p/f occur in the same place as k/t", it's expected more p/f words means less k/t words. So it's not an evidence easy to find I guess.

Indeed. Just as lower-case "b" is less common on head lines than in normal text, "because there are more words with that exotic letter "B" there, and and that means less "b" words".  

Just from observations 1-3 above, one could conjecture that p/f are special letters that stand for rare sounds from other languages (like one may use "è" or "å" or "ø" or "ü" or "qi" in English); or maybe letters that have the same sounds as other Voynichese glyphs, but are used on some words, not as replacements of k/t but just to mark those words as "foreign" (like Romans used "k", "y", and "ph" instead of "ch", "i", and "f" to mark words of Gerek origin; or Germans use "c" and "q" for similar purpose, and the Japanese use katakana instead of hiragana).  And then one would argue that the first word of each parag was usually "foreign", and that ~5% of the words on parag head lines in Herbal and Stars were "foreign", and about 2% of the words elsewhere were "foreign" -- and thus would often have one of those "exotic" letters.

But observation 4 above contradicts this theory.  If "foreign" words were indeed more common in the parag head lines because the semantic structure of a parag required it (like, the names and nicknames of the plants had to be in "Greek" while the explanation of uses etc. was in "Latin") one should see the prevalence of p/f to spill over into the second line when the first line is reduced to half-width because of plant intrusion.  But we don't see any sign of that.  Whether the head line is long or short, the frequency of p/f is high (~4-6% of all glyphs) all through that line, and  practically zero from the second line on. 

The only viable conclusion I see from observation 4 is that the common use of p/f on head lines is not driven by semantics or pronunciation, but is just decoration.  And, in fact, that the decision to use a p/f on the head line (whatever it stands for) was taken by the Scribe on his own, not determined by the Author; because only the Scribe would know where the head line would end.

Looking at tables is suggestive, but relying on our own visual interpretation of a single table can be misleading.

Quite the contrary: the mere distribution-distance numbers can be quite misleading (as they are on this case).  To understand the situation one must look at the distributions and try to understand why the numbers are what they are.

And if one looks at the previous-glyph distributions one can see that p/f are in fact humongously extremely similar to k/t:
   
All four glyphs have only very small amounts of y, l, e, and Ch; substantial □ (space); a lot of o; and zero of the other glyphs - q, r, a, d, Sh, k, and t.  

The reason why the distance between p/f to k/t/ is "large" is basically that p/f are preceded by space ~40% of the time, whereas for k/t that number is only ~15%; and that shift naturally reduces the frequency of o from ~63% to ~40%, and that of y from ~9% to ~6%.  That is all. This shift may be due to many reasons, including use (1) of puffs -- highlighting the first word of parags.

[quoe]However the right-context distance of k/p is 0.34: it's not that big too, but now it's much higher not only than ...[/quote]

Indeed the next-glyph distributions of p/f and k/t are significantly more different than the prev-glyph ones:
   
But, again, the main differences is that k/t are followed by e ~35% of the time, while p/f are never followed by e; and on the other hand k/t are followed by Ch only ~15% of the time, while p/f are followed by Ch ~50% of the time.  Also the frequency of a is ~27% for k/t but ~17% for p/f.  And that is all. The frequencies of space, d, s,o, y, and sh are practically the same for all four letters.

So p/f indeed look different from k/t/ on the right side, but basically because they cannot be followed by e. It would be very interesting to see this table (not just the rms distances!) with additional rows for ke and te. 

... do not support the idea that p/f replace some random k/t when they happen to be in the top-left corner of the page

It is not "in the top left corner of the page".  The plots by @npcompl33t show that the distribution is extremely trimodal, with very different distributions described by my proposed three uses (1-3); and use (2) of p/f seems to be uniform along the whole head line, except for a gap of 5-10 glyphs at the beginning of the line.  And then it drops to zero abruptly at the line break.

But there are more: @npcompl33t presented some in this thread, based on different statistics than left- and right-context.

The statistics that I remember are about the use of puffs on various parts of the parags.  Of course k/t and p/f are different on that respect.  That is in fact what the "replacement" theory says.

All the best, --stolfi