voynich_prosodic_split_report.txt

VOYNICH MANUSCRIPT — PROSODIC SPLIT ANALYSIS
======================================================================
Exploiting the 2-word alternation to separate rhythm from content.

Test: words at even positions (0, 2, 4, ...) vs odd positions (1, 3, 5, ...)

======================================================================
  KEY FINDINGS
======================================================================

1. THE ALTERNATION IS WEAK AT THE CHARACTER LEVEL

   For the biological section (6681 words, ~28,500 characters):
   - Character frequency deltas between even/odd: all < 0.4%
   - Character entropy: even=4.130 bits, odd=4.161 bits (negligible diff)
   - Bigram entropy: even=6.108 bits, odd=6.182 bits
   - Mean word length: even=4.26, odd=4.28

   The character composition of even vs odd words is essentially identical.
   Characters like 'M', 'd', 'g', 'z' flagged as "prosodic markers" by
   extreme even/odd ratios are ALL low-frequency characters (29-123 counts)
   and are mostly transcription artifacts (variant glyphs, uncertain readings).

   Result: NO character-level prosodic markers exist. The alternation is
   NOT carried by specific characters appearing preferentially at even or
   odd positions.

2. THE ALTERNATION EXISTS AT THE WORD-TYPE LEVEL

   Vocabulary exclusivity:
     Biological:    even-exclusive tokens = 27.7%, odd-exclusive = 29.4%
     Herbal_A:      even-exclusive tokens = 34.9%, odd-exclusive = 35.1%
     Recipe_Stars:  even-exclusive tokens = 35.5%, odd-exclusive = 36.3%

   ~28-36% of word tokens appear EXCLUSIVELY at either even or odd
   positions. This is higher than chance for a vocabulary of this size,
   but the most biased individual words have very low counts (5-24 total).
   No single high-frequency word shows strong position preference.

   Chi-squared tests:
     2-char prefix: 340 (df=343) -- NOT significant
     2-char suffix: 227 (df=210) -- borderline
     Morphological templates: 61 (df=70) -- NOT significant
     Word-pair boundaries: 307 (df=322) -- NOT significant

   The chi-squared tests all land near their degrees of freedom,
   meaning the even/odd distributions are statistically INDISTINGUISHABLE
   at the word level too.

3. THE '4o' PREFIX IS NOT A PROSODIC MARKER

   '4o' prefix frequency across sections:
     Biological:    22.2% of words, even=763 odd=721 (bias=+42, +1.25%)
     Herbal_A:       9.2% of words, even=471 odd=485 (bias=-14, -0.27%)
     Recipe_Stars:  16.7% of words, even=812 odd=828 (bias=-16, -0.33%)

   The '4o' prefix shows NO consistent position preference across sections.
   In biological it slightly prefers even; in the other two it slightly
   prefers odd. The bias magnitude is ~1% -- noise level.

4. AUTOCORRELATION CONFIRMS A REAL BUT SPECIFIC PATTERN

   The autocorrelation analysis reveals the alternation exists, but not
   as a simple even/odd vocabulary split. Instead:

   BIOLOGICAL SECTION -- Features with lag-2 elevation:
     Starts '4o':     lag1=0.009, lag2=0.055, lag3=0.029  +0.036 elevation
     Ends '9':        lag1=0.000, lag2=0.031, lag3=0.031  +0.016 elevation
     Ends 'am'/'an':  lag1=-0.015, lag2=0.042, lag3=0.018  +0.040 elevation
     Contains 'c':    lag1=0.003, lag2=0.062, lag3=0.032  +0.044 elevation

   The pattern: features that mark word ENDINGS (suffix '9', 'am', 'an')
   and word-internal structure ('c' presence) show elevated autocorrelation
   at lag 2, while word BEGINNINGS (first character) do not.

   This means: every other word tends to share suffix/structural features
   with words 2 positions away, but NOT with the immediately adjacent word.

   The word length feature does NOT show lag-2 elevation -- the alternation
   is about morphological composition, not word size.

5. CROSS-SECTION CONSISTENCY

   The lag-2 elevation pattern differs across sections:

   Biological (strongest):
     '4o' start, '9' end, 'am/an' end, 'oh' start, 'c' contains  all YES

   Recipe_Stars (moderate):
     '4o' start, '9' end, 'am/an' end, first char, 'c' contains  YES

   Herbal_A (weakest):
     Only '4o' start and 'oh' start show weak elevation

   The alternation is SECTION-DEPENDENT. Biological has the strongest
   prosodic structure; Herbal_A has almost none.

6. WHAT THE ALTERNATION IS

   The data supports a specific interpretation: the Voynich text alternates
   between two morphological ROLES at a 2-word period. Every other word
   tends to be:

   Position A: More likely to end in 'am', 'an', contain 'c', start with '4o'
   Position B: More likely to end in '9', 'y'

   But these are TENDENCIES, not hard rules. The effect size is 1-4%
   per feature. The alternation is probabilistic, not deterministic.

   This is consistent with:
   - A grammatical pattern (e.g., noun-modifier or subject-verb alternation)
   - A formulaic/liturgical pattern (response structure)
   - A table/list with alternating column types

   It is NOT consistent with:
   - A cipher system that embeds rhythm in specific characters
   - Paragraph marks or spacing conventions
   - Transcription artifacts

7. PROSODIC MARKER ASSESSMENT

   The characters identified at threshold=0.3 as "prosodic markers" were:
     'M'  -- EVA capital (variant glyph), 29 occurrences, prefers odd
     'd'  -- standard EVA character, 26 occurrences, prefers odd
     '6'  -- EVA digit (glyph shape), 21 occurrences, prefers odd
     'g'  -- standard EVA character, 123 occurrences, prefers even
     'z'  -- standard EVA character, 37 occurrences, prefers even
     '%'  -- EVA uncertain reading mark, 23 occurrences, prefers odd
     'A'  -- EVA capital (variant glyph), 163 occurrences, prefers odd

   These are NOT prosodic markers. They are low-frequency characters whose
   apparent position bias is statistical noise. 'A' (163 occurrences) has
   a log2 ratio of only -0.37, meaning odd positions have ~29% more 'A'
   than even -- but on such small counts this is within random fluctuation.

   EVA special characters (',' '(' '%' '$' '+' '*' '!') show no systematic
   even/odd preference. The comma ',' (plant intrusion / uncertain boundary)
   is the most common special char (349 occurrences) with a weak odd bias
   (even=157, odd=192, delta=-0.24%) -- barely above noise.

   Cross-section comparison of markers at threshold=0.3:
     Shared bio+herbal_a: '%', 'd', 'g', 'z'
     Shared bio+recipe:   '6', 'A'
     No character is a prosodic marker in ALL three sections.

======================================================================
  STRIPPED TEXT FILES
======================================================================

  Since no meaningful character-level markers were found, the "content
  only" and "rhythm only" split was done at the WORD level:

  Files saved:
    voynich_bio_full          -- all words, dot-separated (35,185 bytes)
    voynich_bio_content_only  -- '4o' prefix stripped from all words (32,239 bytes)
    voynich_bio_even_only     -- even-position words only (17,560 bytes)
    voynich_bio_odd_only      -- odd-position words only (17,624 bytes)
    voynich_bio_rhythm_only   -- binary stream: 1 if word has '4o', 0 otherwise (6,681 bytes)

  Spectral analysis (frequency analysis tool) running on all five.
  Results appended below when complete.

======================================================================
  DEFINITIVE TRANSITION ANALYSIS
======================================================================

  The lag-1 vs lag-2 conditional probability test confirms the alternation
  is specifically period-2, not simple adjacency clustering:

  BIOLOGICAL (6681 words):
    Feature          Base    Lag1_excess  Lag2_excess  Lag2 dominates?
    ends-9           0.499   +0.0003      +0.0313      YES
    ends-am/an       0.129   -0.0145      +0.0416      YES
    starts-4o        0.222   +0.0091      +0.0552      YES
    contains-c       0.383   +0.0028      +0.0616      YES
    starts-oh        0.048   +0.0403      +0.0435      no

  The key finding: words ending in '9' show ZERO lag-1 excess (+0.0003)
  but SUBSTANTIAL lag-2 excess (+0.0313). This means:
    - A word ending in '9' tells you NOTHING about whether the NEXT word
      ends in '9' (lag-1 is zero)
    - But it PREDICTS that the word 2 positions later will also end in '9'
      with +3.1% above base rate

  This is the textbook signature of period-2 alternation: the immediate
  neighbor is anti-correlated or neutral, but the skip-one neighbor is
  positively correlated.

  HERBAL_A (10430 words):
    Feature          Base    Lag1_excess  Lag2_excess  Lag2 dominates?
    ends-9           0.354   +0.1112      +0.0867      no
    ends-am/an       0.141   +0.0143      +0.0127      no
    starts-4o        0.092   +0.0373      +0.0338      no
    contains-c       0.191   +0.1484      +0.1251      no

  Herbal_A shows CLUSTERING instead of alternation. Lag-1 excess dominates
  for all features. Words ending in '9' strongly predict the NEXT word
  also ends in '9' (+11.1% excess). This is run-based structure, not
  prosodic alternation. The herbal section has different compositional
  rules than the biological section.

  RECIPE_STARS (9803 words):
    Feature          Base    Lag1_excess  Lag2_excess  Lag2 dominates?
    ends-9           0.378   +0.0793      +0.0593      no
    ends-am/an       0.187   -0.0351      +0.0380      YES
    contains-c       0.335   +0.0468      +0.0696      YES
    starts-oh        0.063   +0.0126      +0.0317      YES

  Recipe_Stars is mixed: suffix 'am/an' and 'c'-content show lag-2
  dominance (alternation), while '9'-ending shows lag-1 dominance
  (clustering). The prosodic alternation exists but coexists with
  adjacency patterns.

  INTERPRETATION:
  The biological section has genuine period-2 prosodic alternation,
  concentrated in suffix morphology. The herbal section does not.
  The recipe/stars section has a weaker version of the same pattern.

  This section-dependence suggests the alternation reflects CONTENT
  organization (the biological section may describe paired structures
  or alternating categories) rather than a universal grammatical rule.

======================================================================
  SPECTRAL ANALYSIS RESULTS
======================================================================

  PENDING -- 22 structural analysis processes competing on 8 cores (load avg 23.6).
  The five spectral runs launched but have not completed after 20+ minutes
  due to CPU saturation. Files are saved and ready; run when CPU frees:

    # (local path removed)
    [analysis tool] deep_results/voynich_bio_full
    [analysis tool] deep_results/voynich_bio_content_only
    [analysis tool] deep_results/voynich_bio_even_only
    [analysis tool] deep_results/voynich_bio_odd_only
    [analysis tool] deep_results/voynich_bio_rhythm_only

  Expected: the even/odd split files should each still show harmonic
  structure (the harmonics come from sub-word bigram competition, not
  the word-level alternation). The rhythm-only binary stream should
  show weaker harmonics since it's just 0s and 1s with ~22% base rate.

======================================================================
  AUTOCORRELATION DETAIL DATA
======================================================================

  See: deep_results/voynich_prosodic_autocorr.txt

  Summary table -- lag-2 elevation (lag2 - avg(lag1, lag3)):

  BIOLOGICAL:
    Word length:    -0.048 (NO)
    Starts '4o':    +0.036 (YES)
    Ends '9':       +0.016 (YES)
    Ends 'am/an':   +0.040 (YES)
    Final char:     +0.014 (YES)
    First char:     -0.004 (NO)
    Starts 'oh':    +0.008 (YES)
    Contains 'c':   +0.044 (YES)

  HERBAL_A:
    Word length:    -0.026 (NO)
    Starts '4o':    +0.006 (YES, weak)
    Ends '9':       -0.014 (NO)
    Ends 'am/an':   -0.010 (NO)
    Final char:     -0.015 (NO)
    First char:     -0.013 (NO)
    Starts 'oh':    +0.012 (YES)
    Contains 'c':   -0.014 (NO)

  RECIPE_STARS:
    Word length:    -0.043 (NO)
    Starts '4o':    +0.008 (YES)
    Ends '9':       +0.007 (YES)
    Ends 'am/an':   +0.044 (YES)
    Final char:     -0.000 (NO)
    First char:     +0.008 (YES)
    Starts 'oh':    +0.004 (NO)
    Contains 'c':   +0.023 (YES)

======================================================================
  SHUFFLE PERMUTATION TEST
======================================================================

  Testing whether the lag-2 excess is destroyed by word shuffling
  (1000 random permutations):

  Biological section:
    Feature        Original   Shuffled_mean  Shuffled_std   z-score
    ends-9          +0.0313       -0.0000       0.0123       2.6
    contains-c      +0.0616       +0.0000       0.0125       4.9
    starts-4o       +0.0552       -0.0003       0.0127       4.4
    ends-am/an      +0.0416       +0.0000       0.0119       3.5

  ALL features show z > 2.0. The lag-2 pattern is REAL and POSITIONAL.
  It is destroyed by shuffling, confirming it is a property of word
  ORDER, not vocabulary composition.

  The z-scores range from 2.6 (ends-9) to 4.9 (contains-c).
  At z=4.9, p < 0.000001 — essentially impossible by chance.

======================================================================
  CONCLUSION
======================================================================

  The 2-word prosodic alternation in the Voynich manuscript is REAL but WEAK.

  It manifests as:
  - 1-4% probability shifts in suffix choice (am/an vs 9/y) at period 2
  - Elevated autocorrelation at lag 2 for morphological features
  - No character-level markers -- the alternation is purely at word composition level

  It does NOT manifest as:
  - Characters that encode rhythm (no char shows > 0.5% position bias)
  - Vocabulary segregation (same words appear at both positions)
  - Consistent structure across all sections (Herbal_A barely shows it)

  The alternation is most naturally explained as GRAMMATICAL structure:
  the text has a tendency to alternate between two word-form types,
  analogous to how natural language alternates between content words
  and function words, or between subjects and predicates.

  This is a probabilistic linguistic feature, not a cipher artifact.

  Critical finding: the alternation concentrates in word SUFFIXES, not
  prefixes. The morphological ending of a word (am/an/9/y/oe) carries
  the positional signal. This is consistent with inflectional morphology
  in natural language, where suffixes encode grammatical role.