The Voynich Ninja

Full Version: Atoms and elements of the Voynichese alphabet
You're currently viewing a stripped down version of our content. View the full version with proper formatting.
Pages: 1 2
(17-06-2026, 12:23 AM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view....

The parallel frequency distributions of "ir"/"iir"/"iiir" and "iin"/"iiin"/"iiiin" are real — but they don't require scribal error as explanation. Similar word endings behave similarly throughout the VMS. The same parallel distribution exists for every pair of similar suffixes — "-ol"/"-or"/"-al"/"-ar", "-edy"/"-eedy"/"-ey"/"-eey", "-ain"/"-aiin"/"-air"/"-aiir". In each case, the frequency of one variant predicts the frequency of the other. That is a universal property of the text, not evidence for misreading in one specific case. 

Regarding the ratios: I did not claim exact constant ratios. I claimed that similar words have predictable frequencies — "if it is known that 'chedy' is frequent, it is possible to predict that 'shedy' is also frequently used although less frequently than 'chedy'" (Timm 2014, p. 6). The ratio chedy : shedy (1.2) differs from lchedy : lshedy (2.8) — but the direction is overall that "sh" forms are less frequent than corresponding "ch" forms. The specific ratios vary because there is only an indirect relation between similar word types - similar words are products of the same word family operating in the same production environment -  and because low-frequency words fluctuate more — at 119 and 42 instances, a few extra copying events would shift the ratio substantially.

And yes, co-occurrence is an observation — but it is an observation that constrains the explanation. If similar words co-occur on the same pages (+95% page similarity over frequency-matched controls, as quimqu recently measured for Levenshtein distance 1 pairs), then their frequencies will naturally correlate because they are produced in the same local context. The question is: why do similar words co-occur? That is what the production model answers.

And no, the alleged proportionalities are not a mechanical consequence of the self-citation method. It is the other way around. The self-citation method is the most compact description of all the properties we can observe in the Voynich text.
(17-06-2026, 01:19 AM)Torsten Wrote: You are not allowed to view links. Register or Login to view.The parallel frequency distributions of "ir"/"iir"/"iiir" and "iin"/"iiin"/"iiiin" are real but they don't require scribal error as explanation. Similar word endings behave similarly throughout the VMS. The same parallel distribution exists for every pair of similar suffixes — "-ol"/"-or"/"-al"/"-ar", "-edy"/"-eedy"/"-ey"/"-eey",

No, sorry, the pattern in the frequencies of the -ir/-iin endings is not similar to that seen for other endings.  Take for example the words

    chdy            151     chy            183
  chedy    522  chey    369
  cheedy    61  cheey   188
  cheeedy    0  cheeey   10
  cheeeedy   0  cheeeey   0

(Counts are from my working transcription file; your mileage may vary.) Assuming for the moment that 1 = 3, that is a simple proportionality pattern: the insertion of a specific character (d) in a specific context (between the che^x and y) of a series of similar words gives another series of words with similar relative frequencies.  But in the -i^xr series and the -i^yn series the correspondence is one off: the transformation that "preserves" relative ratios is the replacement of in by r (a distance-2 edit).

That is what I take as evidence that dair is somehow analogous to daiin (not to dain), and daiir is analogous to daiiin.  

And the Scribe misreading sloppy in as r is a simple an plausible explanation  for that pattern. Moreover, if all ir endings are actually iin, the structure of the words becomes significantly simpler.  And other things work better too.
  
Quote:The same parallel distribution exists for ... "-ain"/"-aiin"/"-air"/"-aiir".

No, the distributions for these two series of words are off by one, like those of the suffixes -i^xr and -i^yn:

  air    98  ain   150
  aiir   41  aiin  606
  aiiir   0  aiiin  50
  aiiiir  0  aiiiin  0

Quote:The ratio chedy : shedy (1.2) differs from lchedy : lshedy (2.8) — but the direction is overall that "sh" forms are less frequent than corresponding "ch" forms.

But the ratios are not constant, and they are often reversed:

  chy  185  shy  90
  cho   91  sho 138

So the proper description of the situation is rather the opposite: the frequencies of similar (distance-1) words are generally not proportional.

Quote:The self-citation method is the most compact description of all the properties we can observe in the Voynich text.

This claim is trivially and grossly false.  The observable properties of the VMS text include "word #1 is fachys", "word #2 is ykal", "word #3 is ar", and so on.   The text generated by your method may satisfy at most a couple hundred of these properties, just by chance. (Unless the seed text is the whole VMS and the reset and mutate probabilities are set to zero.) 

What you may be able to claim is that a few bulk statistical properties of the output of your method -- like word frequencies, word-word correlations, etc. -- match those of the VMS text.   But surely you can adapt your method so that those statistics of the output match those of Hamlet.  If not, one can always develop a different random text generator algorithm with that property.  Should we take that as evidence that Hamlet may have been created by such method?

I have seen you or some other "gibberish" advocate justify why your method does not produce certain features of the VMS (like the local deviations from Heaps's law in certain sections but not in others, just found by @bi3mw) by saying that the Author was not an automaton, and those special features could be produced by your method if the random choices made by him were not computer-random but merely human-random.  However, with this "modification", your method could also generate Hamlet...

All the best, --stolfi
(17-06-2026, 06:27 AM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view....

The "off by one" pattern is interesting — but it doesn't require scribal error. At the stroke level, "r" and "n" are both word-final single strokes occupying the same position. Mauro's distributional analysis confirms that "r" and "n" behave similarly in word-final context. The parallel frequency series between "ir"/"iir"/"iiir" and "iin"/"iiin"/"iiiin" reflects the fact that adding "i" strokes before either word-final stroke produces a parallel series — because both endings occupy the same structural position. No misreading required.

Regarding cho (91) / sho (138): even this "reversal" confirms the relationship rather than refuting it. 91 and 138 are the same order of magnitude — compare this to natural language, where visually similar words like "the" and "thy" differ by a factor of 10,000. In the VMS, even the reversed pair differs by a factor of 1.5. And exact constant ratios across all pairs would require a counting mechanism, a scribe tracking frequencies and maintaining precise proportions. The variation in ratios — 1.2 here, 2.8 there, occasionally reversed — is the human signature. The ratios are approximately proportional because the process is approximately consistent. The imprecision is the proof that a human produced the text, not an algorithm.

Moreover, "cho" doesn't exist in isolation as a pair with "sho." It exists in a network and is therefore also similar to "cheo," "chol," "chor," "chy," and more. The frequency of "cho" reflects its entire network position — how many similar words exist, how often each is visible, how often each is used as a source. Isolating one pair and demanding its ratio match another pair's ratio ignores the network that determines all frequencies simultaneously.

Regarding Hamlet: self-citation is not a universal text generator that could produce any text. It makes specific predictions through the feedback loop inherent in the copying process: frequent words are more likely to be selected as copying templates, generating more variants; the existence of more variants increases the probability that members of that word family are selected in subsequent copying events. This feedback loop predicts specific features:

- Frequency-connectivity correlation: the most frequent words accumulate the most similar neighbors.
- Continuous evolution: each new word differs minimally from its source; accumulated modifications produce gradual vocabulary drift — the "chol" to "chedy" gradient.
- Asymmetric vocabulary distribution: "words typical for Currier A exist in Currier B, but not the other way round" (Timm & Schinner 2020, p. 8) — because late-emerging variants could not appear on pages written previously.
- Near-identity of type and token word length distributions: since each new word inherits the length of its source, the type inventory mirrors the token distribution.
- Prefix-suffix independence: prefix and suffix are selected independently from visible sources.
- Spatial clustering: similar words co-occur on the same pages (+95%) and paragraphs (+123%) because they are generated from the same visible sources (quimqu).

Hamlet has none of these properties. Self-citation predicts all of them. The VMS has all of them.
As far as I see these answers, @Jorge_Stolfi constructed his alphabet with counts & calculations, using the relation of two appearance' numbers, compared to a relation of numbers for a second character, to find this second character somehow "being just another edition of the first character", right?

I wouldn't cross that bridge, statistical coincidences may not found a proof this way.


(17-06-2026, 06:27 AM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view.[..]
And the Scribe misreading sloppy in as r is a simple an plausible explanation for that pattern.

Apart from the fact that professional scribes should know where to put a tail onto a dash (taken the very small number of mishaps and corrections in VMS, they knew very well what they are doing),
is this theory based upon the assumption of someone reading and writing from an unprecise draft -- but I find many suggestions here that the VMS was dictated. All r-instead-of-n just being hearing errors now...?
However, no one here can answer the question of 'VMS: dictated or hardcopied?', so that is an assumption with a chance.

(17-06-2026, 06:27 AM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view.Moreover, if all ir endings are actually iin, the structure of the words becomes significantly simpler.  And other things work better too.

ir does not look rather shorter or much easier to write; how can it be a really helpful abbreviation then? This is also a question for all other abbreviation ideas.

Many "leading voices" in this forum discuss about "low entropy", "small character set", "high predictability", "frequent repetition": all of it no bringer of joy into work with VMS.
Why could it be useful to even reduce complexity and do things better then?

Anyway, this post was aimed to collect VMS symbols as parts with unique appearances and clearly distinguishable usages for an alphabet, to gain some common ground. Even "mistakes" may fulfill these expectations, but "abbreviations" not so much. We obviously all see something different in VMS.
I would have expected other discussions, but ok, may it be like this here.
(17-06-2026, 08:08 PM)Stefan Wirtz_2 Wrote: You are not allowed to view links. Register or Login to view.As far as I see these answers, @Jorge_Stolfi constructed his alphabet with counts & calculations, using the relation of two appearance' numbers, compared to a relation of numbers for a second character, to find this second character somehow "being just another edition of the first character", right?

Sorry, I didn't quite understand.  But, indeed, it first occurred to me that the ending ir could be similar to iin when I was developing my Voynichese word model and saw that the frequencies of ir,iir,iiir were roughly proportional to those of iin, iiin, iiiin.  Then I noticed the similarity between the shapes of r and a moderately sloppy in.

Since then I have got a few more bit of evidence that (1) the ending ir is indeed equivalent to iin, and (2) that equivalence is the result of scribal error rather than an intentional feature of the script.  Enough evidence that I now generally assume that equivalence in my analyses.  

But, indeed, I have no proof of either statement.  On the other hand, I know of no evidence or argument that contradicts them...

Quote:professional scribes should know where to put a tail onto a dash (taken the very small number of mishaps and corrections in VMS, they knew very well what they are doing),

I believe there are several incorrect claims here:
  • The VMS Scribe was not a professional scribe.  He knew how to prepare and handle a quill pen, and must have had a neat enough handwriting when writing Latin or vernacular; but he did not know the basic practices that professional scribes use when writing a book, such as scoring straight, horizontal, and equally spaced baselines, and vertical margin lines at uniform distances from the vellum edges.  And the sizes of glyphs and spaces are all over the place.  My guess is that his real profession involved writing, but required only clarity, not beauty.  Something like student, teacher, doctor, accountant, or writing letters for someone who (like old Marci) could not write them himself.  And he obviously had no experience as illustrator.  
  • He did not understand Voynichese at all.   One bit of evidence for this claim is the page where he wrote a text interrupted by a plant as if it was two independent columns of text.  The Author must have taught him the set of valid glyphs, and he must have trained writing them until the Author was satisfied with the result; but he probably did not even know how to pronounce them.  He would not need to.
  • He did make many errors, and left many uncorrected.  This is the most likely explanation for the large number of glyphs and glyph combinations that occur only once.  And for the variations of symbol shapes in You are not allowed to view links. Register or Login to view. (although the latter may be better explained by the BEEP BEEP.)  Note that the ink he used was not waterproof (see the stain on f103r), and did not soak into the vellum (as it would into paper), so many other mistakes may have been corrected by wiping them off with a damp "Q-tip".

Quote:is this theory based upon the assumption of someone reading and writing from an unprecise draft? ... I find many suggestions here that the VMS was dictated. All r-instead-of-n just being hearing errors now...?  However, no one here can answer the question of 'VMS: dictated or hardcopied?'

Very few statements about the VMS are certain.  (Even "written in the early 1400s" is dubious.  While vacating my office at the Univ I just found a block of FORTRAN coding sheets, of those that were given to keypunch operators, that I bought more than 50 years ago.  Like those that Friedman & co must have filled when they transcribed the VMS to punched cards. That's to illustrate the fact that the VMS could easily have been written a century or more after the C14 date of the vellum...)

But my belief is that both alternatives above are true.  The original text was first dictated by someone while the Author wrote it down -- on paper -- in his invented phonetic script or shorthand.  Many years later, in Europe, the VMS we have now was copied from this draft (or from some revised version thereof) onto vellum by a local Scribe. 

The main arguments for this second part are the evidence for the "ignorant scribe" claim above, and the observation that no one would want to write something straight from brain to vellum -- because of the relatively high cost of the latter, and the assumption that errors would be hard to erase and correct (even though that assumption turned out to be false in this case). 

Also, clean-copying is slow and boring work.  Writing one page of draft, with a rough sketch of a figure, could take 10 minutes.  Clean-copying that to vellum and fleshing out the sketch, even with "VMS quality" handwriting and artwork, could take half an hour or more.  The Author may have paid the equivalent of $1 per page for the vellum.  If he could hire a starving student or idle accountant at $0.50 per page, he may have chosen to do that rather than do it himself.

And he may have been unable to do it himself because he had poor handwriting or (like old Marci) had failing eyesight.

Quote:
(17-06-2026, 06:27 AM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view.Moreover, if all ir endings are actually iin, the structure of the words becomes significantly simpler.  And other things work better too.
ir does not look rather shorter or much easier to write; how can it be a really helpful abbreviation then? This is also a question for all other abbreviation ideas.

I believe that m is an intentional abbreviation for iin that the Scribe used when space was tight (like near the end of a line) or when he was feeling lazy.  I don't think ir is an abbreviation for iin, but rather a Scribal error.

By "things would be simpler" I am referring to You are not allowed to view links. Register or Login to view.

  Q^q D^d X^x H^h G^g X^y D^e N^n

where
 
Q = { q }, D = { d, l, r, s }, X = { Ch Sh ee Che She eee },
G = { k, t, p, f, ke, te }, H = { CKh, CTh, CPh, CFh, CKhe, CThe, CPhe, CFhe }
N = { n, in, iin, iiin, ir, iir }
and there are restrictions like q being either 0 or 1, h + g either 0 or 1 (that is, at most one gallows), q+d+e+n <= 3, x+h+y <= 2.  Note that r is the only glyph that appears both in the D set and in the N set.  If ir and iir are "quillos" for iin and iiin, the N set becomes just { n, in, iin, iiin }, and each glyph belongs to exactly one of the sets. 

Quote:Why could it be useful to even reduce complexity and do things better then?

Whether there are or not errors will not depend on what we like it or not...

If there are indeed many errors, attempts to solve the riddle that do not recognize that possibility will necessarily fail.  

Anyone who has tried to transcribe the VMS knows that the Scribe often produced ambiguous glyphs and spaces.  Spaces that may or may not be word breaks, glyphs that are halfway between e and i, or between r and s, or between d and g...  Sometimes transcribers will mark those as ambiguous, with comma instead of period and with [r|s] or similar notations.  But in many cases the transcriber will toss a coin, meaning that he will often get it wrong.  And the existence of those ambiguous forms suggests that many glyphs that do look definitely like r were meant to be s, and so on.  

And if the Scribe can produce such ambiguous forms in the clean copy, why wouldn't the Author write many more ambiguous signs on his draft, which the Scribe would then often get wrong?

In fact, assuming the possibility of numerous errors may even make all sorts of analysis simpler.  For instance, if one suspects that final a may be a scribal error for y, even if only some of the time, it would be prudent to map all final a to y and count them together, rather than count each separately.  Collapsing the two will lose some information that may or may not be important, but keeping them separate will make tables longer, and contaminate both counts with an unknown amount of noise.

All the best, --stolfi
(Yesterday, 03:33 PM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view.Sorry, I didn't quite understand.  But, indeed, it first occurred to me that the ending ir could be similar to iin when I was developing my Voynichese word model and saw that the frequencies of ir,iir,iiir were roughly proportional to those of iin, iiin, iiiin.  Then I noticed the similarity between the shapes of r and a moderately sloppy in.

Since then I have got a few more bit of evidence that (1) the ending ir is indeed equivalent to iin, and (2) that equivalence is the result of scribal error rather than an intentional feature of the script.  Enough evidence that I now generally assume that equivalence in my analyses.  

But, indeed, I have no proof of either statement.  On the other hand, I know of no evidence or argument that contradicts them...

I see three additional problems with your error-hypothesis.

First, identifying errors in an undeciphered writing system assumes we know what the correct form should be. But we can't read the text. Declaring "ir" an error for "iin" requires knowing that "iin" is correct — which requires understanding the writing system, which we don't have. The word grammar you use to identify "ir" as anomalous was derived from the same text you are now "correcting." That is circular: build a grammar from the text, notice some forms don't fit cleanly, declare those forms errors, simplify the grammar, and present the simplified grammar as evidence that the errors were real.

Second, every collapse replaces a rare form with a common one — removing elements with high entropy and replacing them with elements with low entropy. "ir" (607 instances) merged into "iin" (4,109 instances). "m" merged into "iin." Each collapse reduces the number of distinct word types and increases the frequency concentration. If the text is already at the lower boundary of what looks like language — Gheuens showed the functional alphabet reduces to about 13 characters — making it more repetitive by collapsing characters pushes it further from language, not closer.

Third, similar glyphs are a key feature of the VMS writing system, not evidence of errors. Currier wrote about this in 1976 in "The Nature of the Symbols" — the glyphs are built from shared base strokes, and "you can make up almost any of the other letters out of these two symbols"(see You are not allowed to view links. Register or Login to view.). The similarity between "ir" and "iin," between "ch" and "sh," between "k" and "t" is the design of the system. Your own word grammar documents the same structure — similar glyphs occupying the same slots. Removing the similar forms doesn't correct errors — it dismantles the system that produces the text.
Both symbols consist of a line and a curve. The primary difference, as seen in the font examples, is that the curve is connected to the either the top or the bottom of the minim. A brief look at VMs shows that this appears to be valid.

If the two VMs letters were the same, then the connection would be all that mattered. Top or bottom would make no difference, implying that anywhere in the middle would do just as well. All that matters in that case is the connection, not the location of that connection. As it seems that the VMs has no examples of such intermediate connections, the symbols would appear to be two separate glyphs.
(Yesterday, 09:02 PM)Torsten Wrote: You are not allowed to view links. Register or Login to view.First, identifying errors in an undeciphered writing system assumes we know what the correct form should be.

That would be a necessary condition to prove that something is an error.  But, even without it, one can conclude that some quirk is more likely to be an error than an intentional feature.

For instance, there are many instances of e or o whose hole is filled in, turning them into big black dots.  Most people seem to consider the filling of those glyphs as meaningless accidents, and transcribe them as e or o.  But there is no proof of this.  Maybe those black dots are a distinct glyph type...

More generally, every single glyph instance in the VMS is different from all the others, so in theory the Voynichese alphabet could have more than 200'000 letters.  But we generally assume that all those glyph instances are handwritten realizations of only a couple dozen glyph types.  So, when we transcribe the text, we must decide first what are those ideal glyph types, by deciding whether the difference between two glyph instances is meaningless handwriting variation or due to them being distinct glyph types.  While some people my disagree, most seem to agree on the glyph types that are represented by distinct letters in the EVA encoding.  

But all this is mere conjecture, without proof.  Until we decipher the encoding, we will never know for sure whether we got it right.  That is unfortunately how we must operate if we want to do any research at all.

Quote:That is circular: build a grammar from the text, notice some forms don't fit cleanly, declare those forms errors, simplify the grammar, and present the simplified grammar as evidence that the errors were real.

Yes.  This reasoning  assumes that, since the Author created the script from scratch, rather than as an adaptation of an existing script, he probably made it nicely symmetrical.  Like the missionaries who created the You are not allowed to view links. Register or Login to view..  Then the probability that some observed feature of the text is an error is boosted if treating it as such would make the system more symmetrical.

Quote:Second, every collapse replaces a rare form with a common one — removing elements with high entropy and replacing them with elements with low entropy. "ir" (607 instances) merged into "iin" (4,109 instances). "m" merged into "iin."

Yes.  But that is not a valid reason to deny the existence of errors.  Entropy does not mean meaningful information contents; it measures the noise as well as the message. Random spelling errors, non-systematic abbreviations, bogus and missing word spaces --- all these things generally increase the entropy of a text.  

If some X are errors for Y, mapping all X to Y will discard meaningful information as well as that particular form of noise.  But if the estimated fraction of X that are errors is high enough, the benefits of the second effect may compensate for the drawbacks of the former one.

Quote:Each collapse reduces the number of distinct word types and increases the frequency concentration. If the text is already at the lower boundary of what looks like language — Gheuens showed the functional alphabet reduces to about 13 characters — making it more repetitive by collapsing characters pushes it further from language, not closer.

Glyph types are not the same as phonemes.  Most written languages that use the Latin alphabet will encode some phonemes with combinations of two or more letters, or represent multiple phonemes by the same letter.  Consider "sh" in English, or the two phonemes "ò" and "ó" of Italian both written just "o".  

My current best guess for the Voynichese alphabet -- excluding what I believe are abbreviations (m), fancy variants (p, f), non-phonetic or "foreign-sound" symbols (q, v, x), mere handwriting variants or deformations (b, g, j, u), and what I suspect are tone or pitch codes (a, o, y) -- has the following letters and multigraphs: { d l r s } + { k t ke te } + { Ch Sh ee Che She eee } + { CTh CKh CThe CKhe } + { n in iin iiin }; that is, at least 4 + 4 + 6 + 4 + 4 = 22 distinct phonemes.

And that count does not include { a o y } or considers the possibility that these glyphs too may be letter modifiers, in addition to e.  

And that count does not consider also the possibility that some of the 22 "letters" above may be parts of additional digraphs in addition to having phonetic values of their own -- like Italian "g" and "n" occur by themselves with distinct sounds but also occur as the digraph "gn" which stands for a third distinct phoneme.  

And anyway written Voynichese may use the same glyph or digraph to represent two or more phonemes, like the two sounds of "th" in English.  Which is almost certain if the system was invented to be a shorthand system to record speech in real time, rather than a full alphabet for other kinds of writing.

Quote:Third, similar glyphs are a key feature of the VMS writing system, not evidence of errors.

Now it is my turn to note that this is just a hypothesis, not a proven fact....

Quote:The similarity between "ir" and "iin," between "ch" and "sh," between "k" and "t" is the design of the system.

Agreed for Ch/Sh and k/t, but not for ir and iin.  The correct comparison would be between r and n, which, like k and t, have one stroke in common and different choices for the other stroke.  But that is not the case for r and in.  Their similar occurrence patterns breaks the apparent asymmetry in the design of the glyphs.

Quote:Removing the similar forms doesn't correct errors — it dismantles the system that produces the text.

Quite on the contrary: considering final ir as a scribal error for iin removes what would be a big wart of that system.

By the way, my explanation for final ir also implies that some final r that are not preceded by i are likewise quillos for in; so that, for example, any dar may have been meant to be a dain.  But I am not yet willing to correct all final r to in, because my word model says that r (like its cousins d l s) can legitimately occur in word-final position, without a preceding i.

All the best, --stolfi
(Yesterday, 10:07 PM)R. Sale Wrote: You are not allowed to view links. Register or Login to view.Both symbols consist of a line and a curve. The primary difference, as seen in the font examples, is that the curve is connected to the either the top or the bottom of the minim.

You are referring to r and n.  But the proposed equivalence, based on the occurrences in the text, is r = in (sometimes), thus ir = iin (always) and iir = iiin (alwas).  

The justification is that the n is often drawn by the Scribe himself with a round bottom and raised slightly above the baseline, so that the whole n looks like a detached version of the r/s plume.  The theory is that Author must have done the same on his draft, and aditionally his round n would have often touched the preceding i -- and the Scribe would then have misread the two as r.

[attachment=16093]

By the way, I also believe that some (not all!) of the d are quillos for k, due to the same basic process. Namely, the k is often written with short legs so that it has the same height as a d.  The latter,on the other hand, is often drawn with a straight right half (sort of like a & ) and an open bottom.  Then the only difference between the two is whether the left stroke is straight and makes a sharp corner at the top, or curved and joins smoothly with the right (loopy) stroke:

[attachment=16094]

However this confusion would not have happened with t, because of the two loops.

All the best, --stolfi
Pages: 1 2