Voynich Evidence Atlas
Could AI help readers and researchers understand this book? Explore its plants, pools, stars and unfamiliar writing, and the clues that might connect them.
I built this Atlas with the team (the AI coding and research agents I direct) to bring manuscript images, earlier scholarship and measurements together. The writing has measurable patterns. Several simple explanations fail in the forms tested. What the words mean remains unknown.
Start with the illustrated stories below, or explore the full record: 51 findings, tested dead ends and research limits. Checks were performed by AI agents within the project, with no outside review or project-wide correction for multiple testing.
Notice how the writing follows the circles. What relationship did the makers intend between the text and the sky? Select an image to enlarge it.
Explore the manuscript
Start with a page that catches your eye. Each link explains a feature of the writing or pictures and why it matters. The scans illustrate the questions; measurement details are in the linked stories.
-
Folio 1v -
Folio 75r -
Folio 99r -
Yale's scan labelled 101v (part) and 102r
Images: Yale University, Beinecke Rare Book and Manuscript Library, MS 408. Public domain. Image sources and checksums.
Look closer
Read the page layout
Folio 1v in full: writing arranged beside a plant. Notice where each line begins and how the writing fits beside the plant. The linked study measures patterns at line openings.
Writing around the pools
Detail of folio 75r: figures in green pools with writing around them. The linked study asks how neighbouring groups of symbols relate across spaces.
Examine the labels
Detail of folio 99r. Plant-label links remain unresolved.
Enlarge a scan to explore the ink, colours and spacing yourself. These images come from Yale's manuscript photographs.
Where AI could help
The opportunity is to connect clues that are difficult to compare by hand: a root drawn on different pages, a repeated group of symbols, or a convention in another historical book.
Compare the pictures
Image tools could help researchers find recurring shapes across pages and historical collections. A resemblance would give a person something specific to examine, rather than a plant name to accept on trust.
Connect the research
AI can help search and organize transcriptions, studies and corrections. This Atlas keeps observations linked to their sources so another researcher can follow the reasoning and reuse the work.
Test a possible reading
A proposed rule could be applied to pages that were not used to invent it. Predicting an independently recognizable detail there would be a stronger step toward meaning than fitting another familiar example.
These are possible routes to progress, not a solved pipeline. The missing link is still a reliable connection between the writing and what it refers to. Research runs are currently paused; the Atlas preserves the work for readers and future investigation.
Evidence stories
What the status words mean
Each finding is graded by how many independent kinds of check agree with it. A kind of check is called an axis: for example another transcription of the text, code written separately, a measurement from the physical page, or an outside published source. The grades, quoted from the project's filing rules:
| Grade | Meaning |
|---|---|
| TRIANGULATED | agreeing lines on ≥3 axes, at least one leaving the project's own pipeline (impl, physical or external) |
| CORROBORATED | agreeing lines on ≥2 axes |
| SINGLE | 1 axis |
| CONTESTED | a disagreeing line not yet resolved |
Standing means the finding is not at risk, superseded or withdrawn. Other status words are shown as the research record has them. Ids such as s132 (a run) or D-086 (a decision) point into the research record.
-
Manuscript & evidence
Lines have a measurable positional structure
Context image: folio 1v
What to notice. Notice how the writing fits around the plant and where each line begins. The research asks whether a symbol is more likely to appear at a line opening than elsewhere. That helps explain how the book was written, even before we can read it.
Certain written forms occur preferentially at line beginnings. The measured pattern also survives line boundaries detected from scan ink.
Read the evidenceResearch status
Current status TRIANGULATED; standing
-
Manuscript & evidence
Neighbouring word edges are linked
Context image: folio 75r
What to notice. Look at two neighbouring groups of symbols. The study asks whether the end of the first helps predict the start of the next. A relationship would mean the writing rules continue across spaces, rather than stopping at each word.
The last written symbol of one word and the first of the next show ordering dependencies, strongest in Currier B. Some short Latin verse comparisons also reach this range.
Read the evidenceResearch status
Current status TRIANGULATED; standing
-
Manuscript & evidence
Word lengths have a rhythm within lines
Shown here: Folio 103r. Compare the lengths of neighbouring symbol groups in the star-marked text. This is an illustration; the linked result concerns the tested corpus.
The linked story uses folio 1v as a context image.
What to notice. Compare the lengths of neighbouring groups of symbols along a line. The study asks whether long groups tend to sit near other long groups, and short near short. This pattern can constrain how the writing was produced; it does not reveal a translation.
Neighbouring word lengths within lines are related. The tested within-line rhythm is not unique to Voynich: Kalevala verse also reproduces it.
Research status
Current status TRIANGULATED; standing
Includes a correction
Read the evidence -
Manuscript & evidence
Many transcribed spaces have a physical ink counterpart
Shown here: Folio 66r. Look for gaps between groups of ink. These illustrate the spacing question; no measured boundaries are marked here.
The linked story uses folio 1v as a context image.
What to notice. Look for visible gaps between groups of ink. Researchers compare gaps like these with the spaces entered by transcribers. This matters because a computer count of words depends on where those spaces are placed.
Scan measurements agree strongly with certain transcribed word spaces. Ambiguous comma boundaries occupy an intermediate gap range.
Read the evidenceResearch status
Current status TRIANGULATED; standing
-
Manuscript & evidence
Tested plant-label links remain unsupported, with a corrected evidence base
Context image: folio 99r
What to notice. Notice the short inscriptions beside plant parts and containers. If these were repeated plant names, they might appear again beside related drawings. The tested links have not established that connection.
D-086 records a research decision. The code s53 identifies the earlier tests discussed below.
The retained plant-label and repeated-label tests do not support the particular text links they tested. D-086 removes the erroneous s53 text-partner tests; the separate s53 image-matching run is unaffected.
Research status
Current status CORROBORATED; standing
Includes a correction
Read the evidence -
Methods & corrections
Preserving whole lines changes the cross-line repetition conclusion
Not tied to one folio
The old cross-line core-repetition deficit does not persist against the line-preserving comparison on the tested inputs. It no longer discriminates between explanations on that basis.
Research status
P-87 identifies this later check; L-377 is the research note that records it.
Current status Accepted bounded correction in
P-87/L-377; no novelty or decipherment claimIncludes a correction
Read the evidence -
Manuscript & evidence
Seven visual correspondences occupy one physical folio
Folios 102r1, 102r2, 102v1
Shown here: Yale's scan labelled 101v (part) and 102r; it does not show f102v1.
What to notice. This open spread contains many plant parts. Several proposed matches come from panels of the same physical leaf. Counting them as separate pages would make the evidence seem broader than it is.
Seven stronger catalogue correspondences used in the grouped-body proposal all target panels of physical folio 102. They are not seven independent physical units.
Read the evidenceResearch status
Current status Eligibility closed; no scoring; proposal not restarted
-
Methods & corrections
A successful calibration is not a Voynich identification
Not tied to one folio
The code db-S141 identifies a test run on manuscripts with known labels.
db-S141 was accepted for calibration of a graph-matching rule on the labelled Penn LJS 419 and Sloane 4016 galleries. No Voynich page was scored or identified.
Research status
Current status Accepted only for Penn/Sloane calibration; no Voynich use
Includes a correction
Read the evidence -
Methods & corrections
Two parked routes have different outcomes
Not tied to one folio
Fontana and Egerton are other historical manuscripts used for comparison. The codes s132 and s133 identify those comparison runs. A frozen gate is a pass mark set before the test.
Fontana s132 failed its frozen agreement gate. Egerton s133 is process-invalid: its read is preserved and unscored. Neither outcome decodes or disproves a meaning for Voynich.
Research status
Current status
s132frozen gate FAIL;s133PROCESS_INVALID and unscored; both parkedIncludes a correction
Read the evidence
Working interpretation
Could this be a specialized reference book?
This working interpretation connects recurring illustrations, different inscription formats and shared writing patterns. A reference or mnemonic collection concerning natural knowledge is one plausible purpose. No verified reading exists.
Observed patterns
Scan measurements of ink gaps agree with many transcribed word spaces, subject to alignment qualifications. Written forms vary with line position, and neighbouring word lengths show dependence within lines. Kalevala verse also reproduces the tested within-line rhythm. These results do not establish linguistic words or complete entries.
Published catalogue comparisons connect plant motifs across sections. Two stronger correspondences link the roots on f18v and f23r to neighbouring fragments 212 and 213 on f102r2. These are resemblance judgments, not identified species or a demonstrated copying direction.
The retained tests do not establish the proposed plant-label links. Visual recurrence has not supplied a stable written name that carries between the paired contexts.
Working inference
These measurements constrain how the writing was constructed. Taken together, the patterns support coordinated construction more firmly than a particular purpose. A shared writing convention operating at word, line and local section scales could accompany a compilation assembled for consultation. Reference use remains an interpretation, not a finding.
A reader might recognize a root motif beneath a larger plant and again among selected plant parts. That makes reference use plausible. Reusing a drawing template or copying a common exemplar also fits. Calling the recurrence a picture citation adds an unproven function; it does not tell us what the nearby writing means.
Inspect the recorded f18v / f23r correspondences and their limits.
Three explanations remain
Reference or mnemonic notation
Pictures and distinct inscription formats might help recognition and consultation. This requires stable textual referents or technical relations that have not been recovered.
Meaningful language transformed by encoding
A code or production convention might obscure linguistic content. This could serve the same reference purpose, but no fixed key, language or successful unseen reading is established.
A coherent visual compilation with structured pseudotext
Templates and disciplined writing can coexist without recoverable textual reference. This also needs a producer that meets the surviving constraints; failures of particular ciphers do not establish it.
What would change this interpretation
A fixed textual mapping that predicts an independently established referent or technical relation on a genuinely unexposed sheet would connect the writing to content. Compositional linguistic recovery would favour encoding; stable nonlinguistic reference would favour notation. A failed prediction would reject that mapping, not every meaningful explanation.
Corrections, scope and source
Version 2, 2026-10-19.
We cannot identify a medical household, Italian workshop, secrecy motive or single production campaign. No semantic winner is justified. The missing prediction is a condition for reconsideration, not a new experiment or compute assignment.
P-87 identifies a later research check. Fontana and Egerton are other historical manuscripts; s132 and s133 identify the comparison runs. A frozen gate is a pass mark set before the test.
P-87 changes the identical-core cross-line repetition conclusion, not the within-line length rhythm. The seven stronger grouped correspondences share physical folio 102; eligibility closed without scoring. Fontana s132 failed its frozen gate, and Egerton s133 remains process-invalid and unscored.
- Preserving whole lines changes the cross-line repetition conclusion
- Seven visual correspondences occupy one physical folio
- Two parked routes have different outcomes
Reader-friendly selection from the existing synthesis, not a new exhaustive review. The full internal synthesis is not included because its local references and redistribution clearance are not prepared for visitors.
Simone Systems Research; interpretation prepared with AI assistance. Underlying observations, catalogue comparisons and prior work retain attribution on the linked evidence stories. Scientific priority is not established.
Download this dated interpretation and its source fingerprint. This is an interpretation supplement; it does not add a finding or change the nine evidence records.
How the stories connect
Some stories use the same sources; others revise an earlier conclusion. Explore these connections to see how the evidence fits together and exactly which claims a correction changes.
Explore the recorded connections
Shared source
Partial correction
- Preserving whole lines changes the cross-line repetition conclusion corrects part of
F-40(research record, not an atlas story) - Preserving whole lines changes the cross-line repetition conclusion corrects part of
F-41(research record, not an atlas story)
Withdrawn line of evidence
- Tested plant-label links remain unsupported, with a corrected evidence base withdraws one line of evidence from
c-377(research record, not an atlas story)
Explore or reuse the research
This is a selected guide to the manuscript and the research I ran with the team. You can follow an individual story or download the underlying records for your own work.
Selection and research coverage
The research record holds 51 findings. This export includes 7 of them, defers 44 for later review and excludes 0, and tells 9 stories. A deferred finding keeps its status in the research record; it is simply not presented here yet.
All graph findings inventoried from existing graph, not all experiments independently reviewed. Two findings included only as correction context.
Download the records, schema and README, and see whose transcriptions the counts rest on.
File details on this page
Format, size and a short SHA-256 identifier, measured when this page was built.
- NOTICE.txtText, 4.8 KB, sha256 1f2de6f51a2d
- working-interpretation.jsonJSON, 5.1 KB, sha256 541db82a501a