Brain-V: What an Autonomous AI Found in One Day

[Brain-V]

Brain-V: What an Autonomous AI Found in the Voynich Manuscript in One Day
I want to share some results from a statistical analysis project and get the community's feedback. This is not a claimed decipherment. It's a systematic hypothesis-testing programme, and everything — including our errors — is published.

Live dashboard: https://brain-v-beryl.vercel.app
Full hypothesis board: https://brain-v-beryl.vercel.app/hypotheses
Failed approaches catalogue: https://brain-v-beryl.vercel.app/failed

What is Brain-V?
Brain-V is a cognitive architecture that runs a loop: generate a testable hypothesis, run a statistical test against the EVA transliteration (Zandbergen ZL3b, 38,053 words, 226 folios), score the result, update beliefs, repeat. Over 60 cycles it generated 123 hypotheses, eliminated 18, and converged on 53 with confidence above 0.8.
The decryption engine is pure Python — no LLM in the statistical tests. An LLM (Claude) generates hypotheses and evaluates scoring, but the maths is maths. All code is open source.
I want to highlight six findings for discussion. Some may be well-known to this community — I'd appreciate hearing where these overlap with or diverge from existing work.

Finding 1: We caught our own false positive
Our word-level codebook attack initially scored 0.87 against Latin. Latin function words (de, et, quod, medicina) appeared in the output. Looked great.
The score was dishonest. It excluded 64% of the text — every word the codebook couldn't map. Honest score including unmapped words: 0.19.
We fixed it, committed the correction, published both numbers. I raise this because any claimed decipherment that reports accuracy only on words it can map is presenting a misleading number. The honest score must include what you can't explain.

Finding 2: 37.3% vocabulary reduction from suffix stripping
Stripping word-final glyphs (y, n, l, r, and multi-character sequences dy, ey, in, ain, oly) reduced unique vocabulary from 8,261 to 5,181 types. This holds across every section.
Corollary: glyph y drops 7–10 percentage points in frequency across every section after stripping. It's overwhelmingly word-final. We believe it functions as a grammatical marker, not a cipher letter. Previous frequency analyses treating y as part of the cipher alphabet were working with a corrupted frequency distribution.
I'd be interested to hear whether the suffix-stripping approach or the y observation is already established in the literature. We arrived at it through statistical testing but recognise this community has decades of prior work.

Finding 3: A three-layer model
After stripping, the hapax ratio barely moved (70.1% → 70.4%). The suffixes are real morphology, but the stems underneath are still overwhelmingly unique. This suggests at least three layers:

Suffix morphology — grammatical markers attached to stems (validated by 37.3% reduction)
Stem cipher — substitution/transposition on the roots (evidenced by IC consistent with monoalphabetic substitution)
Natural language — medieval Latin or northern Italian (evidenced by Zipf R²=0.91, Latin frequency correlation after stripping)

Every approach in our failed approaches catalogue attacked a single layer. The layered structure may explain why: frequency analysis fails because suffixes corrupt the distribution, codebooks fail because stems are enciphered. You have to strip first, then crack, then read.

Finding 4: One cipher system across the whole manuscript
We derived a stem-level substitution table from the biological section and applied it to every other section. Cross-section transfer loss was below 5%.
This suggests one cipher throughout — section-level statistical differences reflect content variation (labels vs. prose, restricted vs. open vocabulary), not different ciphers.
The Currier A/B split is real (confirmed at 0.99 confidence) but appears to reflect different scribal hands or preferences, not different systems.

Finding 5: The biological section is the optimal attack surface
After suffix stripping, blind frequency substitution on stems:
Biological: 0.02% → 6.81% Latin dict hits (340x improvement), IC = 0.100
Text-only: 0.17% → 3.28% (19x), IC = 0.087
Recipes: 0.37% → 3.09% (8x)
Herbal: 0.21% → 2.32% (11x)

IC of 0.100 after stripping is above natural language range for Latin (0.065–0.075), consistent with monoalphabetic substitution on a restricted repetitive vocabulary. The biological illustrations (figures with tubes, body-connected diagrams) support a medical/anatomical text with repetitive terminology.
Combined with Finding 4: crack the biological section's stems and the solution should generalise.

Finding 6: Stem families
Stem analysis found consistent word families:
qo → qol, qotol, qotain, qopar (one root, five surface forms, same stem-to-Latin mapping with different suffixes)
dal → daly (maps to "ab" + suffix y)
ol → olain (maps to "est" + suffix ain)

This is what inflected natural language looks like under a substitution cipher with grammatical suffixes preserved.
Single-character high-frequency words in the text-only section (d, r, o, l — 9 occurrences each) may be function particles or abbreviations (d for de/dosis, r for recipe, etc.).

What Brain-V has NOT done
It has not deciphered the manuscript. It has not produced a reading of any folio. It has not identified the source language with certainty. It has not recovered the cipher key.
It has narrowed the solution space with statistical evidence and corrected its own errors publicly. The full hypothesis board, including all 18 eliminations, is at https://brain-v-beryl.vercel.app/hypotheses.

Community input welcome
I'm aware that this community has decades of deep expertise. I built Brain-V to test hypotheses systematically — if you have a testable theory, I can run it through the same pipeline and publish the results (pass or fail). The most useful contributions are precise, testable claims rather than general impressions.

Dashboard: https://brain-v-beryl.vercel.app
Hypotheses: https://brain-v-beryl.vercel.app/hypotheses
Failed approaches: https://brain-v-beryl.vercel.app/failed

Built by Ben - https://github.com/BuilderBenv1/brain-v.