FINDINGS FROM THE COMPLETED STUDY · public presentation of [preserved evidence reference]
Zodiac semantic-anchor experiment v001
Completed 2026-09-29 (EDT). Verdict: no defensible semantic anchor; qualified negative evidence for this provisional visual ledger. This report addresses structure, not translation. Browse the complete human-facing Results Explorer.
Question and evidence boundary
Can identifiable zodiac imagery and later readable month notes constrain the meaning of original Voynich labels? The signs and months are independently recognizable, but their original text associations are unknown. A nearby original form could identify a sign, month, figure, star, degree, property, date, instruction, or something else. Similar geometry does not select one interpretation.
The later month words and original Voynich writing are separate chronological and script layers. The month words appear beside/within the central emblems in ordinary Latin-alphabet script; their presence does not show that the original writer knew Latin or wrote them. René Zandbergen's month-name table reports the readings below and regards the language as a Romance variety, favoring northern French. Bowern and Lindemann's 2021 linguistic review calls the month words Occitan. The specific variety is disputed. Neither source checked here securely establishes how many people wrote the month notes or their date. The reading of abbreviated letters in September–November remains editorially qualified by parentheses.
Method and verification
The experiment resumed from the , with Gemini, Virgo and Scorpius reserved as whole-sign holdouts. Jeremy clarified that Neither refers to delegated local-LLM machines, not deterministic host computation; no local LLM was used. The code and outputs are and . The v005 authoritative database matched its saved SHA-256 2d2a1b47cec1a2e6768dbf64bd955726b1f5af37eb1730a7d29efdf85e620668; its SQLite integrity and foreign-key checks passed. All eight Yale TIFF views used by the twelve diagrams matched their v005 source hashes. V006 was still in progress and supplied no labels.
The source-reviewed panel crosswalk uses 1006200=f70v2, 1006201=f70v1, 1006202=f71r, 1006203=f71v/f72r1/f72r2/f72r3, 1006204=f72v3/f72v2, 1006205=f72v1, 1006206=f73r, and 1006207=f73v. Yale's partial/composite filenames are retained; scholarly diagram identities are assigned from visible central imagery and foldout position. An initial erroneous f72v assignment produced an apparent adjacency effect. Its full output is ; all reported results below use the corrected source map. The episode demonstrates why this crosswalk is a consequential review case.
We queried the v005 ledger without changing it. The primary set is its default high-confidence manuscript-leaf flag, inside each approximate diagram annulus (normalized radius 0.28–1.16). A sensitivity set adds WRITING_UNCERTAIN_FORM. The centre is shown in the explorer but excluded from original-writing statistics because later month ink is there; this conservative choice can omit some original central writing. Visual-family IDs are provisional geometry, not accepted glyph identities. Angular sectors are twelve-degree position bins, not words, reading direction or figure IDs. Every assigned high/uncertain observation, including centre and margin records, is available in the and human explorer; the full source pictures expose material omitted by v005 extraction.
The deterministic tests use ten whole-sign units, combining both Aries diagrams and both Taurus diagrams. We compared Jaccard overlap of family sets for adjacent versus nonadjacent signs with 50,000 sign-order permutations; a second test excludes sign pairs sharing the same Yale image. Two split-sign Jaccards were compared with 50,000 random diagram pairs. A nearest-training-sign rule was tested on the three preassigned held-out signs, with 50,000 whole-sign calendar-rank permutations. Ring association used 10,000 within-diagram ring shuffles and a maximum-statistic correction across eligible families. is an additional sensitivity control. The small sample, correlations among tests, approximate regions, and inadequate family/label validation rule out an unqualified scientific acceptance claim. Reported p values are permutation tail areas for these proxies, not probabilities of a translation.
Zodiac/month/position matrix
| Diagram | Yale ID | Depicted sign | Later month / nominal season | Published later reading | Original visual material / position |
|---|
| f70v2 | 1006200 | Pisces, two fish | March / spring | mars | 279 high annular observations; figures, central stars, circular text |
| f70v1 | 1006201 | Aries, dark ram | April / spring | aberil | 146 high; fifteen-figure half sign, concentric labels |
| f71r | 1006202 | Aries, light ram | April / spring | aberil | 81 high; fifteen-figure half sign, concentric labels |
| f71v | 1006203 | Taurus, light bull | May / spring | may | 31 high; fifteen-figure half sign, concentric labels |
| f72r1 | 1006203 | Taurus, red bull | May / spring | may | 81 high; fifteen-figure half sign, concentric labels |
| f72r2 | 1006203 | Gemini, paired figures | June / summer | jong | 83 high; radial figure labels and ring writing |
| f72r3 | 1006203 | Cancer, crustaceans | July / summer | iollet | 56 high; radial figure labels and ring writing |
| f72v3 | 1006204 | Leo, lion | August / summer | augst | 80 high; radial figure labels and ring writing |
| f72v2 | 1006204 | Virgo, woman | September / autumn | septe(m)b(r) | 110 high; radial figure labels and ring writing |
| f72v1 | 1006205 | Libra, scales | October / autumn | octe(m)bre | 143 high; radial figure labels and ring writing |
| f73r | 1006206 | Scorpius, animal emblem | November / autumn | nove(m)bre | 130 high; radial figure labels and ring writing |
| f73v | 1006207 | Sagittarius, crossbow figure | December / winter | decebre | 101 high; radial figure labels and ring writing |
The original-material column counts v005 observations, not complete labels, words or figures. Their exact provisional family IDs, source-native boxes, inner/outer bin and angle are browseable for every row. The sign imagery and later month establish the diagram sequence, not the semantics of any original sequence. The March–December season is a nominal modern grouping for comparison, not an inferred medieval calendar rule. No surviving Capricorn or Aquarius page was invented.
Structural results and negative evidence
The twelve annuli contain 1,321 high-confidence v005 observations in 1,093 provisional families; only 30 families appear on more than one diagram. The high-plus-uncertain set has 3,073 observations in 2,671 families, with 64 families on more than one diagram. This fragmentation and uneven coverage matter: f71v has only 31 high annular observations despite many visible marks. Most apparently sign-unique family IDs are therefore not evidence for sign names. Full writing recall and stable family identity were not established in v005.
| Controlled comparison | Primary high set | High + uncertain | Reading |
|---|
| Adjacent minus nonadjacent mean Jaccard | −0.000415, p=0.596 | +0.001780, p=0.097 | No reliable serial progression |
| Same comparison across different Yale views | −0.001458, p=0.795 | +0.000950, p=0.218 | Shared image does not explain a positive accepted result because none remains |
| Mean same-sign split similarity (Aries/Taurus) | 0.0311, p=0.087 | 0.0187, p=0.141 | Two split signs give weak, inconclusive replication |
| Held-out mean month-rank error | 3.67 steps, p=0.604 | 3.67 steps, p=0.603 | Did not predict held-out progression |
| Maximum ring-family deviation | p=0.567 | p=0.293 | No family-wide ring rule survives this control |
The held-out nearest-family rule predicted Gemini near Taurus (one month step), Virgo near Taurus (four), and Scorpius near Taurus (six). This rule was specified in code after inspecting the training framework, but it was not independently sealed before first execution; the holdout outcome is an exploratory predictive check, not a preregistered confirmation. Removing the ten most common v005 families also leaves no robust serial pattern (primary p=0.864; uncertain sensitivity p=0.119). No family or component here supports a month/sign name reading.
The exposes repeated neutral IDs, co-occurrences and approximate order on every diagram. Recurrent leading IDs with changing nearby IDs exist as spatial patterns, but the bins can join different labels, figures, or circular text and v005 does not establish unit boundaries. No base-plus-changing-component family was promoted to a semantic or morphological claim. Prefix/suffix-like or numerical-looking marks likewise remain visual alternatives. The ring test is a coarse positional probe. Figure-by-figure pairing was not independently source-reviewed for all nymphs, so no valid figure-identity association or degree/star reading can be asserted. These are consequential limits, not a null finding about the manuscript itself.
Answers to the seven questions
- Do original labels/text track zodiac sign? Not demonstrated. The provisional family comparisons and the two repeated-sign checks provide no qualified sign-specific rule; incomplete recovery prevents a manuscript-wide negative claim.
- Do they track month/calendar position? Not demonstrated. Primary serial contrast is negative and the sensitivity contrast does not survive its controls. Sign and month are perfectly paired here, so even a positive effect would not separate them without independent evidence.
- Do they track individual figures or diagram position instead? The source images show figure-associated labels, but individual label-to-figure links have not been validated. A coarse inner/outer ring-family test finds no surviving family-wide positional rule. Person, star, degree and position remain live alternatives.
- Are there recurring base+changing-component families across consecutive diagrams? Spatial co-occurrences recur, but no stable original sequence boundary or visual family identity makes such a family defensible yet.
- Does any structural rule predict held-out diagrams? No qualified rule did. The exploratory nearest-family rule missed the three held-out signs by 1, 4 and 6 month steps, averaging 3.67; sign-rank permutation p=0.604.
- Do later month notes constrain an original sequence strongly enough to be a semantic anchor? No. They constrain image/month association for a later reader, without identifying a particular original Voynich label or its referent.
- What did we learn about the zodiac text? The original diagrams contain extensive radial and figure-associated writing, but the current neutral ledger is too sparse and provisional to determine what those sequences denote. The best supported result is a boundary: imagery and later annotations do not license a translation.
Human review and limits
The Results Explorer has all twelve images, later central annotations, 3,355 saved high/uncertain v005 observation rows assigned to the diagrams, angular-sector inventories, side-by-side comparisons and controlled statistics. Its Review Needed subset contains four consequential cases: composite-panel assignment, month-hand interpretation, sparse original-label coverage, and recurring-family identity. Display previews are derived from exact verified Yale TIFFs; their native boxes and source hashes remain linked. Browser rendering is a presentation check, not scientific acceptance or physical-iPhone validation.
This result preserves a qualified negative: no semantic anchor was established using v005 geometry. It does not establish that the historical writer never encoded sign, month, figure or degree information. Reassessment requires better source-reviewed original-label boundaries, figure associations and visual-family reliability, without importing proposed decipherments into the observation layer.