Voynich Clusters
Expanding Currier's voynich languages beyond A / B
Pages are compared using Jensen–Shannon divergence between transition matrices.
Manuscript overview
Golf is the internal name for the regimes described in the forum analysis. The key counts Golf 2.0 support in this view. Intermediate candidates use only their two-color stripes; select a page for its support.
Divergence matrix
Outliers & distances
Outlier scores
These scores use paragraph text, as the Golf clusters do.
Score definition & corpus
Higher scores indicate greater external distance relative to within-page variation. Golf’s score formula uses the same selected text as the matrix and within-group trees.
Pages and references require ≥60 words and ≥20 per line half. References exclude the focal bifolium. Zero-denominator ratios are omitted. A selected supported page appears below the top 12 when needed.
These descriptive ratios are not significance tests. A small denominator can inflate a score.
Internal variation and nearest neighbors
The Y axis compares alternating line halves within each page. Select a point to see its nearest pages.
Read this plot
Higher points have more divergent line halves. Points farther right have a more distant second-nearest neighbor. Both axes show JS in bits.
All pages remain included, with small samples shown as hollow points. Internal JS reflects both sampling noise and differences between lines. Shorter pages tend to have higher internal JS.
References exclude the selected bifolium, but may share a bifolium with each other. Highlighted neighbors remain at their own chart coordinates, so proximity on this plot is not pairwise JS distance. Click empty space to clear the selection.
Isolation at equal word counts
How this ranking works
Ranges span percentiles from random words and contiguous text windows of the same word count. Ranking uses the lower of the two values. Equal values share a rank. The comparison group supplies the samples, while neighbors are searched across the selected text corpus.
The distance is measured to the second-closest external bifolium. Two nearby pages on the same bifolium count as one neighbor, so a single close pair cannot hide isolation. Each trial excludes both the sampled source bifolium and the selected page’s bifolium. Longer source pages contribute 32 or 64 samples per bifolium and method. Bifolia receive equal weight.
At least five source bifolia are required for a rank. Limited comparisons are shown below ranked pages. Missing results mean no estimate is available. These exploratory percentiles are not significance tests, and low scores can miss real outliers in sparse text.
Methods, sampling & data
Distances & samples
Off-diagonal cells compare full page profiles. The gold diagonal compares alternating text lines within a page. It describes sampling and heterogeneity and is excluded from clustering.
Profiles count within-word transitions in Zattera units, including start/end markers. Each context needs at least three observations on both pages. Missing contexts are not reweighted. Colors show JS divergence in bits. Trees use √JS.
The selected corpus retains every page or panel with included words. Paragraph text uses P-coded loci. All text adds C/R loci. Labels and label-only pages are excluded. Orange italic names indicate fewer than 60 words or fewer than 20 in either line half. Sparse pages can appear artificially close. Source category counts use transcription codes in metadata.
Classifications
RZ codes use René Zandbergen’s marker rules, from the 2026-10-04 snapshot. The +/− q-frequency modifiers are merged for grouping. Metadata retains the full codes.
Orders & gaps
Golf and RZ classifications are fixed. Ward linkage and optimal leaf ordering are fitted separately within each group. Group order minimizes the sum of square roots of median cross-bifolium JS distances between neighboring groups. Golf 2.0 retains its hierarchy; Golf 1.0 keeps 3a and 3b together. RZ permits any class order. Whole-group reversals then minimize page distances across block boundaries. Unassigned pages remain last. This arranges existing groups; it does not establish new branches.
Ward on this pair-dependent, context-weighted distance is exploratory and not exactly Euclidean. Its exact minimum-variance interpretation does not apply.
Physical order groups pages by quire and sheet position, then manuscript order within each bifolium. Section gaps follow transitions without moving pages. A gap can split a bifolium visually. The overview and Golf V2 branch chart start in quire order, with gaps between quires. Quire → bifolium groups bifolia within each quire and keeps gaps only between quires. Its menu groups pages by classification and places gaps between those groups. Pages retain manuscript order within each group. A page with multiple hands stays in a separate mixed-hand group.
Branch depth 0 leaves trees unsplit. Depth 1 shows the root split. Larger depths subdivide each tree. Separators mark hierarchy positions, not evidence for regimes or languages. Class and section gaps are twice the selected gap width.
Exports
Matrix CSV uses manuscript order with the split-half diagonal. JSON retains a zero-diagonal pairwise matrix, the separate split-half diagonal, and trees. SVG contains the current order, gaps and values.
Golf V2
Explore the branches, their manuscript locations, and the features that distinguish them.
Branches across the manuscript
Each text-kind unit appears once, in its working group. Select it for the audit and any candidate alternatives.
How to read the assignments
Full color shows strong support from the page itself or validated bifolium assistance. Diagonal stripes in the group color and a lighter shade show tentative support with unresolved weaknesses or conflicts. A lighter fill with white diagonal stripes shows unverified membership, where the working class remains but the audit does not adequately establish the individual assignment. Unverified does not mean probably wrong.
Intermediate candidates use diagonal stripes in their two candidate colors, without an additional support pattern. Select a page to read its support level. Grey means unassigned. The key counts all units in the selected view, including intermediate candidates within their support levels. These levels come directly from Golf's assignment export and are not probabilities or confidence in the existence of a group.
In the all-kinds view, units sharing a page and group stack within one cell. Different groups occupy separate rows. Dashed tree branches are provisional relationships and ? marks an unresolved parent. Branch lengths are schematic.
Transitions that distinguish the profiles
Read these plots
Page circles use the overview fills. Solid color means strong support, main-color and lighter stripes mean tentative support, and lighter color with white stripes means unverified membership. Two candidate colors mark intermediates. Grey crosses are unassigned or outside the comparison. Diamonds show pooled bifolia or within-page blocks, whose saved assessments remain in their details. They do not have page-level support ratings. Point size follows the smaller context denominator, with a minimum size for legibility.
Pages appear above whole bifolia on matching axes. Rosette uses blocks from one face instead of independent bifolia. Each axis needs five context events, or three for Rosette blocks. Omitted points are counted above each panel and valid zero rates remain visible.
The saved feature pairs are descriptive projections of related classification evidence. They are not independent validation or decision boundaries. The table below uses frozen reference profiles. Page colors use Golf's exported working memberships and support levels. Pooled points retain the saved group assessments. Coordinates and rates are unchanged.
Distinguishing transitions and words
Read these tables
Transition rates are conditional on the preceding glyph. ^ marks word start and $ marks word end. Comparators give each sibling equal weight, so these are not necessarily differences from every sibling separately.
Word rates are occurrences per 1,000 words. TF-IDF selects the examples; the displayed rates are observed frequencies. Rare examples fail the repeated-example screen of three occurrences across two sheets when multiple sheets exist. Repetition on one sheet does not establish independent replication.
LSA associations report positive directions across 14 settings. These are sensitivity checks, not 14 independent tests or decoded meanings. Flagged transitions reverse direction under current assignments.
Golf 2.1
Reviewed memberships from word and joined-line transition profiles.
Branches across the manuscript
Read the assignments
Solid color shows strong support. Main-color and lighter stripes show tentative support. Lighter fills with white stripes show unverified membership. Two candidate colors mark an unresolved boundary. Each unit appears once in its working group.
Golf 2.1 reviews the paragraph assignments. Non-paragraph memberships and their support are carried from Golf 2.0. Branch lengths are schematic; dashed branches indicate provisional relationships.
Distinguishing transitions
Read these plots
Each column shows two transitions selected for their contribution to differences between frozen reference profiles. These are descriptive views, not an independent classification test. Five context events are required on each axis. Valid zero rates are shown.
Circles use the reviewed page memberships and support patterns. Diamonds pool only the units included in this comparison from each bifolium; mixed groups are grey. A pool is not independent of its pages. The Rosette comparison has only one physical sheet.
Transition profiles
Read these tables
Rates are conditional on the preceding Zattera unit. Start and end refer to words or lines as indicated. Comparators give each other profile in the selected comparison equal weight. The table shows the ten largest contributions per representation to their combined JS difference.
Reference memberships are frozen; current assignments color the plots. Reference statistics include their own members and describe the groups. Individual assignment reviews exclude the target bifolium. Neither the feature selection nor the share of divergence is a significance test.