{
  "as_of": "2026-11-03",
  "scope": "Targeted primary-source review, not an exhaustive novelty search or independent rerun of the cited studies. No fabricated English translation.",
  "main_conclusion": "The manuscript remains undeciphered according to its holding institution. There is no established whole-manuscript English translation or validated Voynich decoder to rank as most accurate. Peer-reviewed structural analysis, reproducible code, synthetic-cipher accuracy, and accepted decipherment are different claims.",
  "status_sources": [
    {
      "title": "The Beinecke Cipher (Voynich) Manuscript",
      "url": "https://beinecke.library.yale.edu/beinecke/collections/beinecke-cipher-voynich-manuscript",
      "verified": "Current Yale Library collection description says the script is unidentified, cryptographic approaches have failed, and the text remains undeciphered. Illustration-based subject categories are not translations."
    },
    {
      "title": "Voynich Conference 2026: Call for papers",
      "url": "https://www.um.edu.mt/events/voynich2026/callforpapers/",
      "verified": "Current academic conference description still states that the manuscript has resisted decipherment. A conference invitation to study AI/NLP does not validate a particular solution."
    }
  ],
  "matrix": [
    {
      "id": "voynich2vec",
      "title": "Voynich2Vec: Using FastText Word Embeddings for Voynich Decipherment",
      "authors": "William Merrill and Eli Baum",
      "date": "2018 project; author's blog indexed 2018-04-15",
      "publication_status": "Seminar final project, identified as such by the author repository; no peer-reviewed publication established here",
      "primary_urls": ["https://lambdaviking.com/blog/2018/voynich2vec/", "https://github.com/viking-sudo-rm/voynich2vec"],
      "method": "fastText embeddings; suffix/morphosyntactic visualization, page/topic aggregation, attempted MUSE alignment to known-language spaces.",
      "evidence": "Author reports suggestive morphology, inconclusive topic clustering, and unsuccessful MUSE alignment attributed to the small corpus.",
      "reproducibility": "Public code, transcriptions, model folders and pinned MUSE submodule. Original PDF links currently return404; author blog and repository were directly inspected.",
      "not_demonstrated": "Reliable dictionary meanings, Latin/English translation, or accepted decipherment.",
      "comparison_to_v4": "Closest precursor to the PDF idea. Our controlled English tests and bounded official MUSE comparison diagnose parts of the pipeline; they are not an exact reproduction of Voynich2Vec's actual-Voynich experiment.",
      "external_followup": {
        "title": "Word Embeddings for the Morphosyntactic Analysis of the Voynich Manuscript",
        "authors": "Adrien Schurger-Foy, Faustine Rossi, Alexandre Tkatch",
        "url": "https://www.epfl.ch/labs/mlo/wp-content/uploads/2022/10/crpmlcourse-paper1311.pdf",
        "date_status": "EPFL course-project report hosted under2022/10; publication date/peer-reviewed venue not established",
        "finding": "Extends suffix work to prefixes, uses Latin controls, CBOW rather than skipgram and a distinctness metric instead of relying on t-SNE. Some suffix conclusions differ, including-ody,-or,-ar. Partial follow-up exists, but it validates neither word meanings nor a complete decipherment."
      }
    },
    {
      "id": "hauer_kondrak2016",
      "title": "Decoding Anagrammed Texts Written in an Unknown Language and Script",
      "authors": "Bradley Hauer and Grzegorz Kondrak",
      "date": "2016-04",
      "publication_status": "Transactions of the Association for Computational Linguistics4:75–86",
      "primary_urls": ["https://aclanthology.org/Q16-1006/", "https://aclanthology.org/Q16-1006.pdf"],
      "method": "Source-language identification and decipherment under monoalphabetic substitution plus within-word anagram assumptions.",
      "evidence": "Reports97% language identification over380 languages. Table2 gives93.8% word-token recovery on50 generated500-character ciphertexts in5 languages, assuming the source language is known for the decoding steps; language-model corpora contain9–50million words. Hebrew is proposed for Voynich, not established by a full validated reading.",
      "reproducibility": "Open paper; this review did not establish a maintained complete reproduction package.",
      "not_demonstrated": "93% accuracy on the Voynich manuscript. Its repeated-word metric and cipher task are not our160-type evaluation.",
      "comparison_to_v4": "A stronger precedent for validating an algorithm on known ciphers, but different assumptions and test distribution prohibit a direct accuracy leaderboard."
    },
    {
      "id": "bowern_lindemann2021",
      "title": "The Linguistics of the Voynich Manuscript",
      "authors": "Claire L. Bowern and Luke Lindemann",
      "date": "Online2020-11-11; volume January2021",
      "publication_status": "Annual Review of Linguistics7:285–308; scholarly review",
      "primary_urls": ["https://www.annualreviews.org/content/journals/10.1146/annurev-linguistics-011619-030613", "https://alumniacademy.yale.edu/sites/default/files/2021-07/The%20Linguistics%20of%20the%20Voynich%20Manuscript.pdf"],
      "method": "Synthesis of linguistic structure and decipherment hypotheses.",
      "evidence": "Presents arguments for natural-language structure while explicitly acknowledging unreadable contents and no agreed underlying language.",
      "reproducibility": "Review article rather than a single executable decoder.",
      "not_demonstrated": "Academic consensus on a specific language, key or translation. Peer-reviewed counterarguments exist.",
      "comparison_to_v4": "Supports careful separation of structural observations from semantic recovery. Useful methodological orientation, not a validated translation baseline."
    },
    {
      "id": "gaskell_bowern2022",
      "title": "Gibberish after all? Voynichese is Statistically Similar to Human-Produced Samples of Meaningless Text",
      "authors": "Daniel E. Gaskell and Claire L. Bowern",
      "date": "VOY2022 conference2022-11-30/12-01; proceedings posted2023-01-06",
      "publication_status": "CEUR-WS3313; proceedings state double-blind two-phase peer review",
      "primary_urls": ["https://ceur-ws.org/Vol-3313/paper4.pdf", "https://ceur-ws.org/Vol-3313/", "https://github.com/danielgaskell/voynich"],
      "method": "42 volunteers produced meaningless texts; statistical comparison with Voynich transcriptions and meaningful-language controls.",
      "evidence": "Meaningless human-generated writing can reproduce many statistics formerly treated as evidence of language. Authors note their samples cannot test all higher-level page/quire structure.",
      "reproducibility": "Public Python/R code, anonymized text transcriptions, metrics and extended output.",
      "not_demonstrated": "Proof that Voynich has no meaning.",
      "comparison_to_v4": "A particularly useful open negative-control benchmark. Shuffle controls alone are weaker than realistic meaningless text that retains local structure."
    },
    {
      "id": "naibbe2025",
      "title": "The Naibbe cipher: a substitution cipher that encrypts Latin and Italian as Voynich Manuscript-like ciphertext",
      "authors": "Michael A. Greshko",
      "date": "2025-11-26 online",
      "publication_status": "Cryptologia journal article",
      "primary_urls": ["https://www.tandfonline.com/doi/full/10.1080/01611194.2025.2566408", "https://www.tandfonline.com/doi/pdf/10.1080/01611194.2025.2566408", "https://github.com/greshko/naibbe-cipher", "https://doi.org/10.5281/zenodo.16415087"],
      "method": "Constructs a historically plausible verbose homophonic cipher producing Voynich-like statistics from Latin/Italian plaintext.",
      "evidence": "Known plaintext is recoverable for the invented cipher, and several manuscript statistics are reproduced. This establishes compatibility of a cipher family, not the actual cipher.",
      "reproducibility": "Author provides encryption/decryption Python, reference texts and datasets; additional Excel/data supplement linked.",
      "not_demonstrated": "Decipherment of genuine Voynich text or an English translation.",
      "comparison_to_v4": "Stronger positive control for non-word-preserving/homophonic mechanisms. Our whole-token-to-English-word mapping assumes a different structure."
    },
    {
      "id": "rozanova_temerev2026",
      "title": "A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space: What the Units of Voynichese Are Not",
      "authors": "Liudmila Rozanova and Alexander Temerev",
      "date": "2026-08-17 v1",
      "publication_status": "arXiv preprint; no journal reference found on the inspected record",
      "primary_urls": ["https://arxiv.org/abs/2608.17096", "https://github.com/lrozanova/voynich-units"],
      "method": "Tests glyph/token/space assumptions using matched prose, cipher, pseudo-text controls, quire resampling and image-space measurements.",
      "evidence": "Reports strong boundary structure and weak token-order predictability; challenges simple substitution and stable word-boundary assumptions.",
      "reproducibility": "arXiv ancillary code/data and public repository with figure and calibration scripts. Repository says NumPy/Matplotlib/pandas suffice for figure reproduction; OpenCV is needed to rerun raw image measurements.",
      "not_demonstrated": "Accepted decipherment or independently established universal rejection of all language/cipher possibilities.",
      "comparison_to_v4": "Directly relevant test of assumptions preceding our token-level alignment. Open reproducibility is stronger than a bare claim, but is not external replication."
    },
    {
      "id": "kinnison2026",
      "title": "The Voynich Manuscript’s positional entropy collapse: comparative evidence from languages and ciphers",
      "authors": "Rod Kinnison",
      "date": "2026-09-07 online",
      "publication_status": "Cryptologia journal article; publisher abstract inspected",
      "primary_urls": ["https://www.tandfonline.com/doi/abs/10.1080/01611194.2026.2697318"],
      "method": "Character entropy and positional statistics with folio separation, isolated transcription layers and language/cipher controls.",
      "evidence": "Abstract reports that a verbose substitution cipher reproduces several positional statistics while ordinary-language controls do not.",
      "reproducibility": "Full text was blocked by publisher access response; code/data availability and implementation were not independently verified in this review.",
      "not_demonstrated": "A decoded dictionary, underlying language, accepted translation, or proof that the constructed cipher is the original.",
      "comparison_to_v4": "New structural comparator to our intrinsic entropy checks; its existence does not validate our LLM reranker."
    },
    {
      "id": "zhang_sheng2026",
      "title": "Does the Voynich Manuscript Mean Anything? Calibrated Tests Find Little Trace of Language",
      "authors": "Xin Zhang and Victor S. Sheng",
      "date": "2026-10-25 v1",
      "publication_status": "Recent arXiv preprint; not established peer-reviewed consensus",
      "primary_urls": ["https://arxiv.org/abs/2610.34499", "https://arxiv.org/html/2610.34499v1"],
      "method": "Matched-budget copying/cipher generators, held-out statistics, calibrated decipherment controls and planted-message tests.",
      "evidence": "Reports no readable genuine text under tested schemes; explicitly leaves some cipher families unresolved. Negative tests only apply where control ciphers are readable.",
      "reproducibility": "AppendixF says programs and decision rules will be released with the published version. A released implementation was not established here.",
      "not_demonstrated": "Proof of universal meaninglessness or a translation. Very recent author-reported results are not mature external validation.",
      "comparison_to_v4": "Relevant model for calibrated falsification; broader assumptions and realistic nulls than our pilot. Its unpublished-code status limits current independent verification."
    }
  ],
  "unresolved_leads_not_used_as_verified_result": [
    {
      "title": "Singulion Structure and the Voynich Manuscript",
      "authors": "Colin Layfield and Lisa Fagin Davis",
      "year": 2026,
      "primary_url": "https://journals.openedition.org/digitalmedievalist/2331",
      "doi": "https://doi.org/10.4000/16k0a",
      "limitation": "Publisher page blocked by an anti-bot challenge. Secondary dates conflict (July10 versus October19), so do not assert a precise publication date from this review. An author forum announcement confirms June2026 acceptance. This is a manuscript-order/LSA hypothesis, not a decipherment. Full methodological evaluation was not completed."
    }
  ],
  "comparison_and_validation_interpretation": [
    "Our V4 main result93→87/160 and exploratory93→92 do not establish additive accuracy value on held-out known-English cases. Different task definitions prevent comparison with Hauer's ciphertext word-token percentage.",
    "Our internal independent-code audits verify provenance and arithmetic; they are not outside scholarly peer review or independent-laboratory replication.",
    "No accepted Voynich translation accuracy leaderboard was found; rank methods by what was actually validated, assumptions, released artifacts and reproducibility rather than newest date or confident titles.",
    "A stronger GPU can test larger models, longer completed reasoning and fuller training budgets, but cannot manufacture extra Voynich evidence, a bilingual key, reliable segmentation or ground-truth translation. This is an inference from the task and observed limitations, not a demonstrated hardware scaling law.",
    "Before purchasing additional compute, reproducing open structural/segmentation controls and realistic nulls is more informative than requesting fluent full-manuscript English. A supposed complete English rendering without a falsifiable mapping would be invented interpretation."
  ]
}