Open problems and unresolved contradictions
Honest documentation of the statistical anomalies and methodological limits that the system does not explain. Introduced in v7.5.
Methodological caveat
The Hebrew reading model is one of several models and is unsupported in its current form: so far its readings do not tell the manuscript apart from meaningless control texts (OP15). The following problems remain visible even where they count against the model. Internal coherence does not prove external validity.
| Severity ↓ · Status → | |||||
|---|---|---|---|---|---|
13 of 13 problems
Original claim: the word-length distribution is too narrow for free Semitic morphology (rigid concentration on 4–8 EVA characters). Tested scale-invariantly against real Semitic references (Mishnah, WLC): the Voynich coefficient of variation (0.347) lies within the range of the references (0.326–0.363), not narrower. The 4–8 window mentioned holds fewer tokens (79.69 %) than any equally wide window in the Semitic references. Details, key figures and method: see the word-length validation.
Key figures, tables and method: Word-length validation.
Certain EVA glyphs occur statistically only in certain word positions (slot grammar after Tiltman/Stolfi). The binding was first quantified in v9.9.7, formally modeled and calibrated against two Hebrew reference corpora (Mishnah, WLC). At glyph level, 49 % of the positional entropy is predictable from glyph identity; at word level, the binding is 2.4 to 2.5 times that of both references, confirmed by a second, independent measure (restrictiveness of a formal slot model). The formal model OP2 asked for is therefore in place; the word-level finding is robust and confirmed twice. Key figures, tables and method: see the slot-grammar audit.
Working hypothesis:
The expectation that the slot grammar is mainly an affix phenomenon covered by R41 is refuted at word level: conditional stripping of the confirmed affixes lowers the binding by only about 13 %. Also refuted is the expectation that Mishnaic Hebrew, the language the reading model assumes, shows a higher binding of its own so that the gap shrinks. The Mishnah measurement lies practically on the WLC value. At ROOT LEVEL, however, the finding is NOT decided: mutual information and restrictiveness contradict each other (1.86 times bound vs. practically equal); for short strings the second measure no longer separates, so it does not contradict the MI; it is blind at this point. On the status of ch-: ch- is NOT an edge allograph (3.2 % line-initial vs. 71.3 % for p-), and its bare counterpart occurs less often than for established R41 shells, which argues against shell status; ch- and sh- share 57.7 % of their remainders. The question cannot be settled distributionally; it needs minimal pairs chX/X with a difference in meaning in the lexicon.Key figures, tables and method: Slot-grammar audit.
Second-order conditional entropy was re-measured unit-matched against the WLC reference text (Voynich H-layer, 32,646 tokens from 202 folios; WLC 301,837 Hebrew tokens from 39 books). The main comparison is the consonantal skeleton, defined identically on both sides as the character sequence after removing all vowel signs, on the Voynich side after applying the mapping table. Voynichese lies at 2.276 bit (95 % interval 2.240 to 2.310), the Hebrew reference corpus at 3.739 bit (3.701 to 3.766); the gap is 1.46 bit at comparable mean word length (3.86 vs. 3.92 consonants). The gap is stable against the choice of analysis unit and lies between 1.3 and 1.6 bit in every matched convention. Inventory size, type repetition, immediate word repetition and the compound units of the mapping table together explain about 0.2 bit.
Working hypothesis:
The direct measurement does not support tracing the deviation back to the positional predictability of the slot system. The compression is sequential, not positional: knowing the position lowers the entropy in the manuscript by 0.32 bit (H1 3.307 to H_pos 2.983), knowing the preceding character by 1.03 bit; in the Hebrew reference corpus the same values are 0.12 and 0.32 bit. Removing the prefixes and suffixes confirmed by the segmentation lowers the root H2 to 2.000 bit and widens the gap to 1.64 bit instead of narrowing it: the morphological affix layer does not carry the compression, it lies in the root material itself. Re-encoding Hebrew reference text into EVA through the mapping table reaches 3.239 bit and does not reproduce the manuscript values either, at a coverage of 59.7 percent of the characters. The finding is homogeneous across language-A anchor folios and the remaining folios (2.303 vs. 2.227 bit). What remains as an explanation are mechanisms that constrain word-internal character sequences independently of morphology and vocalization.Key figures, table and method: Entropy audit.
No independent Hebraist has blind-tested the system so far: no specialist unfamiliar with the decipherment system has tried to read Voynich text with it. That would be the strongest form of external validation. Separate from this is the coherence blind test under OP15 (30 passages, three raters from outside the project): it rates the glosses of translated passages against control texts, not the Hebrew reading itself.
For Hebraists and Aramaists who wish to carry out an independent check: the complete character mapping and the 10 frozen type-I anchors are ready.
The decipherment-system document is available as a Markdown export.
The exclusivity measurement is defined and deterministically reproducible, but not yet ready to be closed. After collapsing readings (N=1000, seed 42), the TopGap median over analyzable tokens is 31 and the exclusivity rate 73 %, below the closing threshold of 80 %. The rule coverage of phase 2 (15/60 rules) includes every rule that can be checked per token. Independently of this, the coverage rate of the reading generator (69 %) remains open.
Working hypothesis:
The remaining ambiguity is not a broad spread but a single class: prefix vs. root consonant (72 % of the genuine cases), concentrated on a few high-frequency function stems (ol/al, ar/or, aiin, ky, ot). This class cannot in principle be resolved on the isolated token; it requires a contextual CUT (phase 3) on ordered paragraph sequences, analogous to phase 2b. Three-part closing condition: (a) exclusivity rate over analyzable tokens of at least 80 % with a TopGap median above 20, measured after morpheme-chain collapse and with a fixed seed, N at least 1000; (b) full phase-2 rule coverage for the rules classified as per-token capable; (c) coverage rate reported separately, not folded into the exclusivity figure. Of 60 rules, 15 are per-token capable (token-morphological, not discourse/register/structure/meta) and all 15 are formalized in the generator (R1, R2, R3, R4, R8, R28, R40, R41, R45, R50, R59, R61, R65, R66, D1-D2); (b) is therefore met. R17 (shol medial vs. sheol line-final) and R30 (sho- + noun) cannot be formalized per token: R17 depends on line position and language-A/B context, which the isolated-token generator architecturally lacks; the same kind of gap that keeps R44 out of phase 2b; the only lexicon-tagged attestation of R30 (shodaiin = shod + daiin) is a compound of two independent lexicon words rather than a prefix+root structure, and mechanical sho- stripping would also wrongly split the far more frequent shol/sheol words. R44, R46, R52a remain untested borderline cases. The coverage gap of 31 % is a lexicon and generator problem and is not counted as ambiguity.25/25 folios show FOLIUM text-freeness (R60 corollary 3, v8.7: ★★★★★). The 25-folio threshold was passed with f32r, f32v, f33v, f34r, f34v. The basic question remains open: deliberate pharmacological encoding or a purely layout-related phenomenon?
Quire E opens with f33r (Mandragora officinarum, humanoid) as a zoomorphic root depiction. f34v was first read as a zoomorphic Bryonia (Bryonia pair hypothesis); the f34r audit, however, shows a fibrous, tuberous root without zoomorphic shaping and an umbellifer (Seseli libanotis). Question: does f33r remain an isolated zoomorphic case in Quire E, or does the pattern continue in f35r/f35v and further folios?
Three places defined different prefix inventories: R41 named four slot classes (d-; o-/qo-; l-/k- shared; op-/of-), the GRAMMAR_PREFIX table ten entries (some of them mechanical:false), and the mechanical segmentation its own list of nine entries. Consequences (app side, fixed since v9.9.3): k- was never read as a prefix although an R41 slot and a table entry existed (1,145–1,352 word-initial k- attestations, depending on the transcriber); op-/of- was never recognized as a unit but split into o- + p-, or not recognized at all for f-; p-, y-, t- were counted as R41 shells although they have no slot in R41; qok-/qod- were defined only in the segmenter; two independent check paths within the app (the automatically generated rule attestations vs. the /api/gibberish/analyze endpoint) could give opposite verdicts for the same word (ytoldy, otochor, oqotar). In addition, a corpus re-audit showed: l- and k- are strictly sequential, not shared slots (l- before k- 127:2, qo- before k- 310:0, k-/t- complementary: 0 stacked attestations), and op-/of- is the outermost wrapper, not the innermost shell (positional binding 11.4 % vs. bare p- 70.9 %), so not an edge allograph.
Working hypothesis:
Implemented (v9.9.3): GRAMMAR_PREFIXES (grammar.data.js) is the single, hand-maintained app source with class assignment and R41 hierarchy position (r41Slot). The mechanical strip list, the R41 check (checkR41) and the grammar display table are derived entirely from it, with no parallel list any more. k- and op-/of- are now mechanically active (k- guarded: only with an established remainder or a following Currier gallows, since k-/ch- before a vowel is often a stem allograph: lkeedy 39× vs. leedy 0× vs. lcheedy 9×); qok-/qod- were dropped as entries and resolve as qo-+k-/qo-+d-. The three original sub-questions are answered empirically: (1) p- remains a separate OP-A question, clearly separated from op-/of-; (2) k- is segmentable on its own but lexically guarded, not unconditionally free; (3) op-/of- is distributed across the corpus (343 attestations / 125 folios, concentrated in the recipe section f103–f116) instead of being limited to Quire H. Both app check paths named in the problem (the automatically generated rule attestations and /api/gibberish/analyze) therefore derive their prefix inventory from the same source.Certain glyphs are strongly bound to line position rather than to lexical content: 70.5 % of all `-m` tokens are line-final, 71.0 % of all `p-` tokens line-initial. `y-` as the first glyph is a pure position-1 step (pos. 1 = 0.141 → pos. 2 = 0.030 → middle = 0.038). The lexicon-based edge-artifact check confirms the pattern at word level: for several prefix forms the bare stem is edge-neutral, while the same form with a prefix is edge-extreme and almost exclusively line-initial: `chor` 8 % edge vs. `ychor` 100 %, `qotchor` 86 %, `pchor` 67 %, `tchor` 61 %, `ochor` 67 %; `shol` 10 % vs. `dshol` 80 %, `pshol` 86 %, `tshol` 83 %; `shor` 20 % vs. `dshor` 67 %; `cheol` 9 % vs. `ycheol` 93 %, `pcheol` 67 %; `chol`/`or` 14 % vs. `ychol` 92 %. Mirrored for the final `m` allograph: `ram`/`sam`/`oram` 100 % final, `dam` 66 % final (59 of 89 attestations), for `dam`, a high-frequency content lemma, without a deletion verdict. Additional quantitative evidence from the R41 rule audit (v9.9.2): the line-initial share per initial glyph cleanly separates edge allographs from genuine prefixes: p 71.0 % (factor 4.8), t 41.7 % (2.8), y 38.7 % (2.6), f 31.0 % (2.1) vs. o 15.1 % (1.0) and l 6.7 % (0.5) at a baseline of 14.8 % (H-layer, 37,097 tokens, 5,489 lines). Across all p-/f-initial R41 violation types attested in the corpus, 116 of 135 occurrences are line-initial (85.9 %; label/marginalia loci excluded), and 58 of 104 p-initial types have a corpus-attested counterpart without p-, usually orders of magnitude more frequent (pokeey/okeey 168x, pokain/okain 139x, pokar/okar 127x, potol/otol 87x). The OP-A normalization is therefore not only relevant to the lexicon but a mandatory first step of any prefix segmentation.
Working hypothesis:
Edge allographs (initial Currier gallows, final `m`, `y-` prothesis) form an orthographic layer above the linguistics, not an independent lexical signal. They are modeled as a normalization step before the morphological reading; no token receives a lexicon entry merely because of its edge position. The decisive closing criterion is positional confinement (the share of a form’s occurrences that are line-initial): `ychor` 16/16, `qotchor` 12/14, almost exclusively line-initial. A genuine, freely combining proclitic would regularly occur line-medially too; confinement is the signature of an edge allograph. Remediation policy: an edge-extreme prefix form with an edge-neutral stem is not deleted but merged with or reassigned to its stem (merge/reassign). Deletion is only considered when there is no stem relation and the number of attestations stays low; high-frequency content lemmas (e.g. `dam` n=89) never receive a deletion verdict. The label/marginalia register (`L.`/`M.` loci) and single-token lines are excluded from the check, since they are trivially at the edge and carry no allograph signal of their own. Live evidence: see the edge-artifact check section, currently 18 merge/reassign candidates based on the current lexicon and corpus data. The affected lexicon entries are marked as OP-A candidates in the lexicon itself (non-destructive: existing translations stay until OP-A is decided).Content×content 1-edit neighbors, at 1.86 %, lie 3.20× above the content baseline (0.58 %; 597 affected pairs, after excluding the top 20 formula tokens), e.g. `kshy/kchy`, `okay/oky`, `choty/chotey`, `chody/schody`, `chor/cthor`, `dol/dolo`. Falsification test (M1): lexicon classification of the 597 pairs (a_DISTINCT 17.8 %, a_DERIVED 8.2 %, b_VARIANT 1.3 %, b_UNLEX 72.7 %) plus a satellite test for positional binding.
Working hypothesis:
The satellite test refutes the self-citation hypothesis: rare near-twins are not bound to their hub (adjacency binding of rare tokens 1.1 % vs. frequent ones 5.9 %; only 3 % of the rare multi-occurrence tokens ≥ 50 % bound vs. a 1 % shuffle baseline, 0 % fully bound, so no satellite population). The neighbor similarity is morphological adjacency (proclitic chains such as `shol`→`dshol`, `chor`→`ychor`, `chy`→`kchy`, and root echo), not a generative self-citation artifact. Restriction after comparison with OP-A: this does NOT hold for the initially confined proclitic forms (`ychor`, `dshol`, `qotchor`, `tchor`, `dshor` etc.): positional confinement (see OP-A) marks them as edge-allograph candidates for merge/reassign, not as lexically freely combining morphemes, and thus as an OP-A case rather than evidence for the substrate. For non-confined root-echo/adjacency pairs the statement stands: typical of language, it supports the substrate rather than refuting it. The real remainder (only 26 % of the pairs resolve into two distinct readings, including single-gallows swap pairs such as `kshy`/`kchy`, `chor`/`cthor`) shifts to OP5 (exclusivity) as a control metric.The same reading pipeline (lexicon, CUT decomposition, word-by-word glosses) reads meaningless Markov-2 text from Voynich glyphs and Latin Naibbe ciphertext almost as well as the manuscript: confident CUT readings 43.6 % and 43.3 % vs. 45.2 %, lexicon hits 57.7 % and 50.7 % vs. 59.0 %, with overlapping confidence intervals. In the blind test (30 passages, three raters, rule fixed in advance), the translated Voynich passages do not read as more coherent than translated Markov text (mean 2.20 vs. 2.30; p = 0.68). The pipeline thus measures similarity to the Voynich word surface, not meaning. Details: see the model comparison.
The manuscript has a hapax share of 0.669 and a Zipf α of 1.011 (34,000-token samples). Verbose ciphers match H1, H2 and positional binding with every plaintext language tested, but reach only a hapax share of 0.36–0.46 and α 0.90–0.95. Hebrew through the project mapping stays at α 0.80–0.86; the pseudo-word generator matches the hapax share but lies at α 0.41. No model reaches both values at once. Unencrypted natural language comes close to both: in the Zipf validation, Dante’s Commedia reaches α 0.995 and a hapax share of 0.624 (38,000 tokens). Details: see the model comparison and the Zipf validation.
Hebrew and Aramaic text, transferred into EVA with the project’s own mapping, reaches H2 2.87–3.46 bit (manuscript 2.12) and a positional binding of 0.19–0.41 (manuscript 0.70), across 22 variants of the open table decisions. The mapping is also incomplete and not reversible: ז and צ as well as several vowel signs are missing, a strict table encoding matches only 47 of 434 lexicon entries, and the readings do not have a Hebrew letter distribution (ב, ג, ז, צ never occur in the lexicon). Read backwards, the readings inherit the Voynich structure (H2 2.25 vs. 3.06–3.22 for real Hebrew with an equally small vocabulary). Details: see the model comparison.