Brain-V

Brain-V Dashboard

An autonomous cognitive architecture attempting to decipher the Voynich Manuscript (Beinecke MS 408, c. 1404-1438) through statistical analysis, hypothesis generation, and iterative testing.

Active Hypotheses
162
92 above 0.8
Eliminated
24
won't retest
Parked
7
need more data
Beliefs
51
Cycles Run
64
Latest Surprise
0.0219
2026-04-15

Corpus Profile

Words: 38,053
Unique: 8,261
Glyphs: 25
Folios: 226
Glyph entropy: 3.8627 bits
Zipf exponent: 0.8946
Zipf R²: 0.9084
Hapax ratio: 70.1%
Strongest positive finding2026-04-16

Three mutually-reinforcing findings: _.oii encodes psychoactive plant content

Brain-V has a cluster of three hypotheses that lock together: (1) _.oii fires 4.84x more on plant folios than non-plant herbal folios (held-out 5-fold validated), (2) within plant folios, _.oii fires 8.14x more on psychoactive plants than non-psychoactive (pre-registered, confirmed), (3) no volvelle variant tested can reproduce the plant-folio enrichment. Together these three results constitute Brain-V's first defensible evidence that EVA encodes botanical content visible in the illustrations.

H-BV-PLANT-010.75

_.oii marks plant-identified folios

4.84x enrichment vs non-plant herbal baseline
100% 5-fold cross-validation stability; z = 3.88
n: 115 plant folios vs 14 non-plant herbal baseline
H-BV-PSYCHOACTIVE-010.80

_.oii further enriched on psychoactive plants

8.14x ratio (psycho 2.80% vs non 0.34%)
pre-registered one-tailed Welch p = 0.046, Cohen d = 2.63, Mann-Whitney U p = 0.000015
n: 4 psychoactive (Paris quadrifolia, Cannabis, Rhododendron, Nymphaea caerulea) vs 112 non-psychoactive
PRE-REGISTERED · thresholds met: p < 0.05, d >= 0.4
H-BV-VOLVELLE-OII-010.92

No volvelle architecture reproduces the plant enrichment

Four architectures tested; max null 2.80x vs real 5.04x
v1 simple 0.00x; v2 section 26-root max 2.80x z=23; v3 section 500-root max 2.66x z=7.7; v4a folio-level 26-root max 1.82x z=12.3; v4b folio-level 200-root max 1.68x z=11.7. All four empirical p=0.
n: 1,400+ null runs across 4 variants

Structural argument

Under the volvelle hypothesis, plant folios and non-plant herbal folios draw from the SAME section cartridge. Expected enrichment is ~1.0x by construction. Across 500 null runs spanning three architectures, max observed is 2.80x. Real is 5.04x. The finding is structurally incompatible with section-level volvelle mechanics, independent of rate calibration.

Remaining escape hatches

  • Non-volvelle mechanical process (Markov chain, HMM) - untested
  • Sherwood's plant-ID labels correlated with some other folio-level text property - untested
  • Counter-intuitive: folio-level cartridges produce LOWER enrichment than section-level because independence averages signal to 1.0x - the volvelle family is structurally ruled out

Strategic note

This is Brain-V's strongest positive finding. Unlike H-BV-VOWEL-01 (which fell to H-BV-VOLVELLE-01 trivially), H-BV-PLANT-01 survives three volvelle variants at p<0.01 each. Combined with H-BV-PSYCHOACTIVE-01 as a pre-registered sub-class confirmation, Brain-V has evidence that EVA encodes botanical content at a resolution deeper than mechanism signature alone.

Headline finding2026-04-15

Coverage as a decipherment metric is dead

1,300 phonotactically plausible nonsense skeletons drawn from the corpus's own bigram-Markov distribution, size-matched to Voynichese word-length distribution. 20 independent trials.

SystemCoverageΔ vs null floor
Brady Syriac (1,334, claimed)86.90%+3.34pp
Null nonsense (1,300)83.56%
Schechter Latin (4,063)82.81%-0.75pp
Hebrew medieval (1,300)57.90%-25.66pp
Null size
1,300
Null coverage
83.56% ± 1.03pp
Trials
20

Implication. Nonsense skeletons that match corpus word-length and bigram distributions achieve 83.6% coverage with sigma=1pp over 20 trials. Any abjad-reducible lexicon of 1,000-4,000 entries will hit ~80% on the Voynich corpus by virtue of length and bigram structure alone. Brady sits 3.3pp above the null floor; Schechter sits below it. Only Hebrew is statistically distinct from random, and it is 25pp worse. Coverage-based 'decipherment' claims require a z-score against a corpus-derived null, not a random-alphabet null.

Negative result (published)2026-04-15

Vowel-section decode FAILS held-out validation

4 folios (f101r, f89r2 pharma; f78r, f82r bio) removed entirely from training. Rules and naive-Bayes classifier re-derived on the remaining corpus. Every held-out token predicted from vowel pattern alone. Scored against ground-truth section label.

Target: pharmaceutical (prevalence 44.3%)

methodprecrecF1
always-predict0.4431.0000.614
rule table0.8270.1460.248
naive Bayes0.6730.2290.341

Signal does not survive: best F1 below always-predict baseline.

Target: biological (prevalence 55.7%)

methodprecrecF1
always-predict0.5571.0000.716
rule table0.5680.7820.658
naive Bayes0.6550.6000.626

Signal does not survive: best F1 below always-predict baseline.

Per-folio NB hit rate

f101r (true: pharmaceutical, n=226)22.6%
f89r2 (true: pharmaceutical, n=233)23.2%
f78r (true: biological, n=293)61.4%
f82r (true: biological, n=285)58.6%

Held-out pharmaceutical folios mis-classified as biological at ~75%+ rate. Multi-class NB accuracy 43.6% vs majority baseline 55.7%.

Interpretation. The 100% agreement on _.eo -> pharmaceutical in training (8 skeletons) was an in-sample artefact. On held-out pharmaceutical folios, the classifier confidently predicts biological for ~75% of tokens. What survives: rule precision 0.827 on pharma when the rule fires — a real but sparse signal covering 40% of tokens. H-BV-VOWEL-CODE-01 demoted 0.75 -> 0.35. Aggregate chi-square coupling (H-BV-VOWEL-01) is unaffected; what failed is per-token prediction, not distribution-level coupling.

Why this is published. Brain-V publishes this negative result instead of suppressing it. The coverage-game critique (H-BV-NULL-01) still stands; the vowel layer still has aggregate signal; but the per-token decoding path is not the right frontier. Future work: sparse high-precision mappings, non-vowel structural features (glyph-role combos, line-position interactions), or image-label targets instead of section-metadata targets.

Positive finding2026-04-15

EVA vowels encode section-linked information in 79% of testable skeleton groups

For every consonant skeleton with >=3 vowel variants and >=100 tokens, chi-square test of variant distribution across the 8 manuscript sections. Critical value at p<0.01 from df.

Skeletons tested
70
Significant at p<0.01
55 / 70 (78.6%)
Headline skeleton 'kdy' vs chance
5.15× over p=0.01

Case: skeleton 'kdy' — Brady's chedy vs chody case (§3.10). Same consonant skeleton, different vowel pattern.

varianttotalbiologicalrecipesherbalzodiac
chedy501181199624
chdy1401740523
okedy1164131223
okeedy1083647154
chody88023430

chi² = 262.2, df = 28, critical at p<0.01 = 50.9 5.15× over threshold

Top-5 skeletons by vowel-section coupling strength

wkdy [qokeedy,qokedy,qokeody]8.51× over p=0.01
l [ol,al,l]5.79× over p=0.01
kdy [chedy,chdy,okedy]5.15× over p=0.01
sdy [shedy,sheedy,shody]4.99× over p=0.01
kl [chol,cheol,okal]4.88× over p=0.01

Implication. EVA vowels are not padding. Vowel choice within a fixed consonant frame correlates with section (herbal vs biological vs recipes vs zodiac) strongly enough to reject the independence null at p<0.01 in 55 of 70 testable skeleton groups. This is a language-independent structural property of Voynichese. Any future decipherment that treats EVA vowels as noise, or as free positional slots, is discarding information that the manuscript demonstrably encodes.

Three-Lexicon Comparisonnew

2026-04-15

Three independent decipherment lexicons (Latin, Syriac, Hebrew) run through Brain-V's honest pipeline against the full EVA corpus. All three fail the shuffle test on word-order syntax.

LexiconEntriesCoverageconn→content Δboth-matched Δ
Schechter Latin/Occitan4,06382.8%+0.0000
Brady Syriac (proxy, 71 terms)7148.2%-0.0098+0.0149
Hebrew medieval medical1,30057.9%-0.0144+0.0281

Currier B > A across all three lexicons

Schechter Latin+8.21%
Brady Syriac (proxy)+3.92%
Hebrew medieval+3.07%

Three independent methodologies, same direction. Currier A structurally resists lexical matching.

H-BRADY-02 confirmed: gallows are paragraph markers

EVA 'p' line-initial5.0× (pred 5.4×)
EVA 't' line-initial2.8× (pred 3.2×)
bench gallows cth/ckh/cph< 0.5×

Brain-V's first independent verification of an external Voynich structural claim.

Verdict: Three independent lexicons from three language families produce (a) varying but substantial coverage (48-83%), (b) zero connector->content word-order signal under shuffle test, (c) reproducible Currier B>A asymmetry (+3-8pp). Lexicon-based methods establish thematic clustering, not decipherment. Currier A appears structurally distinct from B in a way that resists lexical matching regardless of source language.

Leading Hypotheses

View all
H001
0.99

The manuscript's herbal section uses a combination of substitution and transposition ciphers, which would explain the higher entropy levels compared to other sections.

cipher64 tests run
H020
0.99

The Voynich text encodes a natural language using a null-cipher or homophones, where multiple glyphs map to the same plaintext character, which would explain the high hapax ratio (70.1%) and lower glyph entropy (3.86 bits) relative to expected natural language entropy while preserving Zipf-like structure.

cipher43 tests run
H021
0.99

The Currier A/B split reflects two different scribal hands encoding the same underlying language with different but related cipher alphabets, such that glyph-level bigram transition matrices in A and B sections are structurally isomorphic under a permutation mapping.

cipher43 tests run
H022
0.99

The zodiac and astronomical sections use a systematically different word-order encoding than herbal and recipes sections, reflecting a positional transposition cipher layer applied on top of substitution, detectable as reduced local bigram predictability at section boundaries relative to within-section transitions.

structural43 tests run
H023
0.99

The high hapax ratio (70.1%) is partially artifactual, caused by consistent scribal abbreviation or word-compounding conventions where morphological suffixes are concatenated inconsistently, such that word-final glyph sequences 'y', 'n', 'l', 'r' function as detachable morphological markers — splitting words on these terminals would reduce unique vocabulary by at least 25%.

language43 tests run

Current Beliefs

View all
0.57The manuscript contains meaningful content, not random glyphs
0.60The underlying language is most likely Latin or Italian
0.84Different sections may use different encoding methods
0.31The cipher is performable by hand with period tools
0.91Positional glyph constraints reflect cipher structure not grammar
0.94The herbal section is the best starting point for decipherment

How Brain-V Works

Perceive

Parse the EVA transliteration (Zandbergen ZL3b). Compute glyph frequencies, entropy, Zipf fit, positional constraints, Currier A/B statistics across all 226 folios.

Predict

Generate testable hypotheses about the manuscript's cipher, language, and structure. Each hypothesis specifies the exact statistical test that would confirm or deny it.

Score

Run each test against the corpus. Update confidence scores. Eliminate hypotheses that fail. Promote those that pass. Log everything on-chain via AgentProof.