Thirty Names per Sign: the Zodiac Ring Labels of the Voynich Manuscript as Two-Letter Sigla, and a Falsification Pipeline for Degree-Name Cribs

Preprint archived on Zenodo: doi:10.5281/zenodo.21761982

Author: Vitaly Averyanov (independent researcher). AI assistance disclosure: as in Paper 1.

Abstract

The circular "zodiac" diagrams of the Voynich Manuscript (ff. 70v-73v) carry short labels, one per nymph, about thirty per sign. Building on the workshop-cipher model of the companion paper, we show that these labels are not running text but a distinct register with its own statistics, and that they are best described as two-letter sigla of names drawn from a closed per-sign list of roughly twenty to twenty-five distinct entries. Four independent lines of evidence converge: collapsing the labels by the text's homophone classes or morphological cores does not increase the (near-zero) overlap between rings, so the rings carry genuinely different names (no internal crib); a capacity analysis shows each label encodes essentially two plaintext letters, so full names and numerals do not fit; simulation of candidate name sources matches the observed ring diversity only for a "distinctive-word" profile of ~20-25 types per thirty; and pairs of labels within a ring share exactly one of their two parts above chance, the signature of an independently drawn prefix+suffix encoding - a signature absent from the pharmaceutical labels, which name unique objects. We then present a reproducible three-stage pipeline for testing whether any external per-degree name list is the source of these sigla - a profile filter, a positional match under the blind-replicated degree alignment, and an anachronism control - and report a systematic negative program: seven classes of published degree-lists (degree-images, syllabic calendars, saints' calendars, planetary monomoiriai, paranatellonta, star catalogues, and herbal/pharmacopoeial name repertoires) are rejected under pre-registered criteria. One candidate (the sanctoral of Lleida in a 1524 print) passed a Bonferroni-corrected positional test and an independent transliteration replication, yet was shown by an anachronism decomposition to be a selection artifact: its signal lived entirely on feasts that postdate the manuscript. We argue that the zodiac labels are decipherable in principle but require an external, order-aligned list not yet identified, and we make the falsification apparatus available so that future candidates can be adjudicated rather than merely asserted.

Keywords: Voynich manuscript, cipher, sigla, pre-registration, false positive, zodiac, onomastics

1. Introduction

Whatever else the Voynich Manuscript contains, its circular "zodiac" diagrams (ff. 70v-73v) offer the single most tractable target in the whole codex. Each sign is drawn as concentric rings of nymphs, and beside almost every nymph stands a short label - about thirty per sign. These labels have three properties that running Voynichese lacks: they are bounded (a fixed count per ring), they are positional (arranged around a circle in a definite order), and, if the companion paper is right that the text is a workshop cipher, they are the cipher run on isolated short strings rather than on continuous prose. Prior work established that the labels behave like an inventory of names rather than like text - high type-token ratio, and almost no vocabulary shared between rings (inter-ring Jaccard ≈ 0.005) - and that under a blindly replicated alignment the attribute of each nymph encodes the sex of her degree in al-Biruni's table §457 (shift 23, p = 0.0088). If any part of the manuscript can be read against an external referent, it is here.

That very tractability is a trap. A ring of about thirty short items will "sort of" match a great many thirty-long lists - saints, degree-images, stars, calendar syllables - and the literature of Voynich "solutions" is largely a record of such matches asserted rather than tested. Our methodological stance is therefore the opposite of pattern-hunting: we pre-register the source profile before looking at any candidate, correct for the number of candidates tested, add a period/anachronism control for any source that postdates the manuscript, and report every negative. The point of the exercise is not to announce a reading but to build an instrument that can tell a real source from a plausible coincidence.

This paper makes three contributions. First (§3), we characterize the labels themselves and show, by four independent tests, that they are two-letter sigla of closed per-sign name lists of roughly twenty to twenty-five distinct entries. Second (§4-5), we present a reusable three-stage pipeline for testing whether an external per-degree name list is their source, and apply it to seven classes of published lists, all of which are rejected under pre-registered criteria. Third (§6), we document an instructive false positive - a Lleida sanctoral that passed multiplicity correction and replication yet was exposed as a selection artifact by an anachronism decomposition - and draw the general lesson that such a control is mandatory. We conclude (§7-8) that the labels are decipherable in principle but await an external list that is both period-appropriate and order-alignable to the rings, and we release the falsification apparatus so that the next candidate can be adjudicated.

2. Data

• ZL3b-n Lz loci (with the &-prefix fix: without it 270/300 labels are lost). Robustness on

IT2a-n. Nine signs with exactly 30-label sequences (light-half-first convention, test #13).

• Parse cross-witness. The label inventory does not depend on our own IVTFF handling: an

independent implementation (Yoshida, seeton/Voynich-public, label_locality) reports 883 L* records, 1021 known tokens and 747 types on the interlinear source, and our parser reproduces those figures exactly - 883 of 883 loci in common, no record with a differing token count, and all ten locus classes (L0, La, Lc, Lf, Ln, Lp, Ls, Lt, Lx, Lz) agreeing on record, token and type counts.

• Blind-replicated degree alignment: attribute of nymph = sex of degree per al-Biruni §457,

shift 23 (p = 0.0088, blind round 2); used here only as a fixed alignment, not re-tested.

• On re-examination we read this blind attribute signal, though replicated, as a coarse

light-half/dark-half partition rather than a fine per-degree sex table: a 15+15 model reaches M = 60/78 (p = 0.024) on the same blind vector, at least matching §457's M = 52/78, and the attribute lies in contiguous arcs (runs test p = 0.0013; a single mid-ring boundary fits 46/56 of Aries+Taurus, p < 2×10⁻⁵). The partition is, however, local: two further blind rounds extending the headdress attribute to the six consecutive signs Cancer-Sagittarius (124 and 119 coded figures, contrast-enhanced crops) found no blocking that survives correction across the six signs (0/6 signs; combined runs p = 0.16 and 0.86), so the half-arc structure belongs to the barrel signs and does not license a manuscript-wide lunar or degree reading. A pre-registered specificity control fixes what the alignment does and does not establish: against 20,000 random tables of the same shape - identical block lengths, order permuted - 25.6% reach M >= 52. The blind attribute is thus blocked in a way that agrees with some partition of the ring, but the data do not single out al-Biruni's table in particular; the Latin recension transmitted by John of Seville, the version actually current in fifteenth-century Europe, does worse (46/78, its profile peaking at shift 20). We therefore treat §457 as one tabular realization of a local partition and use the alignment only to fix ring order; the crib rejections below are in any case shift- and key-invariant.

3. Labels are two-letter sigla of closed name lists

Four independent tests converge on a single description of the labels.

No internal crib. If the rings looked non-overlapping only because the same name was spelled with different homophones, then collapsing the label vocabularies by the text's own homophone classes - or by shared morphological cores - should raise the overlap between rings. It does not: the mean inter-ring Jaccard after collapsing is no higher than under random partitions of the same coarseness (p = 0.82 for twin classes, 0.91 for cores). The rings carry genuinely different names, and there is no cheap internal shortcut to reading them; an external referent is required.

Capacity. Using the reconstructed tables, running-text words fit one-letter encodings almost exactly (mean 4.45 glyphs vs a modelled 4.43 for a single letter), whereas labels fit two-letter encodings (5.37 vs a modelled 5.15), and not one of the 344 labels exceeds the two-letter support. A label therefore carries essentially two plaintext letters. Full Latin names (vir, mulier, rex) and Roman numerals do not fit in a single label; numerals are independently excluded by three separate tests. A ring of thirty labels encodes about sixty plaintext letters - not a list of words but a list of two-letter sigla.

Source profile. Simulating the cipher over candidate name sources and comparing the resulting ring statistics to the observed values (in-ring repeat rate 4.1%, inter-ring Jaccard 0.02, TTR 0.805) discriminates sharply among source types. A "first word" source, in which every degree is named by a generic head-word (vir, homo, mulier), produces far too many collisions; an "all distinct" source produces too few. Only a "distinctive word" profile, with roughly twenty to twenty-five distinct types per thirty degrees, reproduces the observed statistics. This fixes the shape of the target source without identifying it.

Two-part signature. Because the two letters of a siglum are enciphered by two independent table draws, a name repeated within a ring should yield label pairs that share exactly one of their two parts (the same prefix under a different suffix, or vice versa) far more often than chance. They do: pairs sharing exactly one part number 455 against a permutation null of 355 (p < 0.001), with both the prefix-only and suffix-only excesses individually significant. The ratio of one-part to both-part sharing implies about eight effective variants per part - richer than the six-table reconstruction, consistent with a lost original apparatus. The control is decisive: pharmaceutical labels, which name unique objects and so contain no repeated names, show no such signature (p = 0.6). The two-part signature is thus tied specifically to name collisions within the zodiac lists.

The o- marker. Sixty-seven percent of labels are o-initial, against about twenty-six percent expected from the tables. This excess is achievable "for free": twenty of the twenty-three letters have an o-initial homophone available, so a scribe who prefers it marks a word as belonging to the name class without spending any encoding capacity. This explains an old observation - that the o- marker is inseparable from the slot grammar and behaves unlike a detachable determiner - as a stylistic choice among homophones rather than an added morpheme.

4. A falsification pipeline for degree-name cribs

• Stage 1 (profile): simulate the cipher over the candidate source; require in-ring repeat

rate M1 and inter-ring Jaccard M2 (and TTR M3) within 95% intervals.

• Stage 2 (positional match): pairs of degrees sharing a two-letter siglum must map, under

shift 23, onto label pairs sharing a part; permutation p; Bonferroni over all candidates tested (running k).

• Stage 3 (anachronism control): decompose a positive signal into pairs supported by the

epoch source vs pairs unique to a later witness; a genuine source cannot live on the latter.

• Openness rule: "discovery" = raw p < 0.005 AND replication on an independent transliteration.

5. A systematic negative program (seven source classes)

Applying the pipeline to every class of published per-degree or per-plant name list yields, so far, only rejections. Table 1 summarizes; each row is a pre-registered test whose script and seed are in the repository.

Two rows deserve comment. The star-catalogue rejection is structural, not incidental: a thirty-degree belt segment crosses only a handful of constellations, so any construct built on constellation names yields five to nine distinct sigla per sign against the required twenty to twenty-five - no rule for choosing the star can repair this. The herbal and pharmacopoeial repertoires fail for a different structural reason: they cannot be aligned to the manuscript at all. The Voynich herbal section is not ordered by name - the first glyph of successive herbal paragraphs is non-descending in only 0.559 of adjacent pairs, against 0.577 expected at random, where an alphabetical herbal would give ≈ 1.0 - so no positional match is possible; and the names are not individually extractable, because verbose capacity fragments a name across several words, the name is diluted into a recipe formula, and the medieval name is not the Linnaean binomial we would test against. This holds regardless of the source language: Persian and Arabic pharmacopoeias (al-Biruni's Saydana equates some 4500 plant names across 27 languages) are richer repertoires and, being the same al-Biruni whose §457 gave our degree alignment, are strategically interesting - but they are alphabetically ordered and so face the same alignment barrier. Their shorter phytonyms ease the capacity barrier while worsening the false-match risk that motivates this entire methodology.

6. An instructive false positive: Lleida 1524

The most valuable episode in the negative program is a candidate that almost survived. Among nineteen diocesan saints' calendars (Usuarium; missals and breviaries, manuscript and printed, thirteenth to sixteenth century), those of the Aragonese-Catalan zone - a region we had adopted from the customary Occitan reading of the marginal month names - cluster at the top of the positional test, and the calendar of Lleida (Illerda) in a 1524 print reaches S = 0.251 against a null of 0.174 with p = 0.0020, passing Bonferroni correction at the twenty-three candidates then tested. It replicates on the independent Takahashi transliteration (p = 0.0040), and a pre-registered hold-out restricted to non-adjacent day-pairs - which removes any artifact of the feria-inheritance rule - passes on both transliterations (p = 0.0043 and 0.0073). By every standard we had built, this was a signal: a period-plausible, region-plausible source, surviving multiplicity correction, replication, and a hold-out.

It is nonetheless spurious, and one further control shows why. The 1524 print postdates the manuscript by roughly a century; a fifteenth-century scribe could not have known feasts added to the calendar after his death. We therefore obtained the actual Lleida sanctoral of the manuscript's own period - the customary RC.0031 of the Old Cathedral, dated 1400-1457, with its 1451 breviary calendar - and tested it as a confirmatory candidate. It fails (p = 0.151). Decomposing the 1524 signal settles the matter: partitioning its predicted pairs into those also predicted by the fifteenth-century sanctoral and those unique to the 1524 print, the signal concentrates entirely on the anachronistic pairs (S = 0.378, p = 0.003) and is null on the epoch-supported pairs (p = 0.77). A source that did not yet exist cannot be the source. The Lleida signal is a selection artifact: the best of a correlated cluster, landing by chance on feasts that a later editor happened to add.

The lesson generalizes. Multiplicity correction and replication guard against two failure modes but not against a third - a candidate whose apparent fit rests on features absent from the period. For any source that postdates the artifact, a period or anachronism control is not optional. We know of no prior Voynich "match" that has applied one.

A second layer of the same lesson concerns how the candidate pool was framed in the first place. Our regional prior came from the customary Occitan attribution of the marginal month names. That attribution traces to a single informal consultation in 1997 pointing to one Toulon document, was never collated against its proposed source, and is not the current consensus: the diagnostic forms aberil, jong and octembre are read by most present opinion as northern French, with a Franco-Swiss alternative, and the month-name hand is in any case a later addition by neither a Voynich scribe nor the f116v annotator. Nor is the northern reading itself on firm documentary ground: in the localised and dated charter corpora we checked, octembre is well attested and northern, but aberil, jong and iollet do not occur at all, so both attributions rest on forms whose readings have never been audited. The false positive therefore arose inside a candidate pool chosen by an unexamined premise. Pre-registration disciplines the test; it does not audit the prior that decided what was worth testing.

7. Discussion

The zodiac labels are the manuscript's most tractable content precisely because they are bounded, positional, and class-marked, and our results sharpen rather than dissolve their promise. They are decipherable in principle: two-letter sigla of a real, closed list of degree names, with a homophonic apparatus whose internal structure we can measure even without the key. What they are not is readable from the inside. Reading them requires an external list that is simultaneously of the right period and region and order-alignable to the rings - and the rings are not internally ordered by name, so alignment must come from the source's own known order (as with the blind §457 degree alignment) rather than from the manuscript.

Prior and concurrent observations. Two items outside the peer-reviewed record bear directly on these results, and we state them rather than leave them for a reader to find. First, the concentration of o-initial labels is not new with us: it was reported on the voynich.ninja forum in July 2025 (thread 4829), where seventy-seven percent of 299 zodiac labels were counted as o-initial and within-ring duplicates were tabulated. What we add is not the observation but its status - the marker is embedded in the slot grammar rather than being a detachable determiner (§3), and it is cheap precisely because most letters have an o-initial homophone available. Second, Honeycutt (2026) independently reports structural differentiation in the same rings - including container-holding formats confined to particular signs, which agrees with our blind result that barrels occur only in Pisces, Aries and Taurus - and holds that a "30-per-sign decanic count" is falsified in favour of a "10-sign architecture" with a 15/30/29 count distribution. That claim concerns figures per folio, not labels per sign, and on inspection it agrees with our data rather than contradicting them: the surviving zodiac covers ten signs; Aries and Taurus are each split across two folios carrying fifteen labels apiece, which recombine to thirty; Gemini (f72r2) carries twenty-nine and is for that reason outside our nine-ring set. "Thirty names per sign" is a statement about recombined per-sign label sequences, and it is exact for the nine rings used here.

Seen this way, the negative program is progress, not failure. Before it, any thirty-long list was a candidate "solution"; after it, the space of surviving hypotheses is small and explicit: a lost local list of the right region (which no published corpus can supply), or the per-degree star descriptions of the Liber Hermetis, which remain to be tested when a machine-readable text of Gundel's edition becomes available. Everything else in the published record - degree-images, syllabic calendars, saints' calendars, monomoiriai, paranatellonta, constellation catalogues, and herbal or pharmacopoeial name repertoires - has been excluded under stated criteria, and the same criteria will adjudicate the next candidate in minutes.

Finally, the label register connects directly to the mechanism of the companion paper. The labels are the same verbose apparatus applied to isolated short strings, which is why they appear in their "full table form": a canonical prefix-core-suffix parse in sixty-three percent of labels against forty-three percent in the running text, where serial reuse and graphotactic shortening deform the words. The zodiac labels are, in effect, the cipher caught standing still - the cleanest window we have onto its per-word behaviour.

This two-register split is measurable directly, and it locates the manuscript's discriminative content. Treated as corpora, the running body behaves like the workshop generator - cross-page Jaccard 0.068, type-token ratio 0.198, second-order entropy h2 2.18, against 0.090 / 0.107 / 2.10 for a Naibbe-class generator on Latin - while the labels stand apart on every axis (Jaccard 0.006, TTR 0.728, h2 2.53). The content that distinguishes one ring from another therefore sits in the labels, not in the surrounding prose: the body reads as procedure, the names as payload. The body's type-token ratio is about twice the generator's, so it is not pure filler; but it is far closer to generated text than the labels are, which is why the names, and not the paragraphs, are where an external referent must be sought.

8. Conclusion

The zodiac ring labels of the Voynich Manuscript are two-letter sigla of closed, per-sign name lists. They can be read only against an external, period-appropriate, order-aligned list of degree names, which is not yet identified; seven published classes are excluded here under pre-registered controls. The apparatus we provide - profile, positional match, and anachronism decomposition - is offered as a standard for adjudicating future crib proposals, so that the next candidate is tested rather than believed.

Data and code availability / Acknowledgements / References

• Repo (shared with Paper 1). Corpora: calendar_*_usuarium.tsv, calendar_lleida_consueta15c,

plant_ids_tvg, almagest_v61, cisiojanus_grotefend14c.

• Al-Biruni §457/458 (Wright 1934); Davis 2020 (hands); Usuarium (ELTE); The Voynich Garden;

Pelling 2009 (cisiojanus); Greshko 2025 (cipher, Paper 1); Firmicus (Kroll-Skutsch); Fité i Llevot 2016, Miscel·lània Litúrgica Catalana 23: 135-180 (Lleida consueta); VizieR V/61 (Almagest); Yoshida, K. (2026), Voynich Research: Multiscale Structural Constraints and Executable Generative Models, https://github.com/seeton/Voynich-public (CC BY 4.0; label_locality dataset used as the independent parse cross-witness in §2).