A reproducible, conservative research toolkit and technical journal for studying the Voynich Manuscript (Beinecke MS 408).

This repository does not present a final decipherment. Its goal is to organize structural hypotheses, research routes, analysis scripts, and human-in-the-loop visual-review tools so that ideas can be tested without turning inference into "translation."

The main research line — "what is Voynichese?" (routes R43–R71) — is now closed:

• Working model. Voynichese behaves as a morphologically rich, syntactically thin token system: a restricted generative process (templatic operators qo-/ok-/ot- plus matrix borders) over a weak prose-topic layer, with the text statistically decoupled from the images.
• Capstone (R62 / R67). A local, content-free generator reproduces 13 of 14 statistical signatures of the corpus — effectively 14/14 once the last residue (LAAFU) is shown to be page layout, not content. Meaning is therefore not necessary to explain the statistics — which is not the same as proving the text is meaningless. The statistical line is exhausted.
• Remaining uncertainty is provenance/material, not statistics. Current priors: content-free generator ≈ 70% · constructed system/notation ≈ 22% · recoverable cipher ≈ 8%. Only an external key or a held-out decode could move the meaning question.
• No decipherment or translation is claimed anywhere in this arc; every result ships with an explicit guardrail.

Long-form summary: resumo_voynich_codex.md. Consolidated closure report: docs/summaries/relatorio_R43_R61_natureza_do_voyniches.md.

Each visible token is modeled as a layered form rather than a single substituted symbol:

[visible token] = mode operator + nucleus/template + border value

The a/o axis carries most of the textual structure; the r/l axis is weaker and may be visual/diagrammatic.

git clone [email protected]:Pantani/voynich.git
cd voynich
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r voynich-codex-project/requirements.txt

cd voynich-codex-project
PYTHONPATH=. python -m pytest -q   # 650+ tests

Python ≥ 3.10 is required. Core dependencies: requests, pandas, matplotlib, pytest.

Download the public text corpus used by the experiments:

python scripts/download_sources.py   # fetches ZL3b-n.txt from voynich.nu

Download public folio images from the manifests:

python scripts/download_images.py

Run a single route script:

PYTHONPATH=. python scripts/analyze_form_collocations.py

voynich/
├── README.md                   ← you are here
├── resumo_voynich_codex.md     ← long-form executive summary
├── CONTRIBUTING.md · NOTICE.md · CITATION.cff · LICENSE
└── voynich-codex-project/      ← all research code, data, and docs
    ├── scripts/        ← 80+ route scripts (one per step; flat — do not move)
    ├── src/voynichkit/ ← Python package: clean API (corpus, types, constants)
    ├── data/
    │   ├── raw/            ← ZL3b-n.txt (main EVA corpus)
    │   ├── annotations/    ← human-fill sheets (pipeline gates)
    │   ├── derived/        ← computed CSVs (never edit by hand)
    │   ├── transcriptions/ ← small manual EVA seeds
    │   └── manifests/      ← image manifests & source lists
    ├── docs/
    │   ├── research/    ← one report per route (rota_1 … rota_71)
    │   ├── foundations/ ← hypotheses, glossary, sources, attack log
    │   ├── summaries/   ← executive summary, route map, closure report
    │   └── tools/       ← static HTML review/calibration tools
    ├── images/         ← downloaded folios (raw/, git-ignored) + derived crops
    ├── tests/          ← 650+ pytest tests
    └── project_state.json ← machine-readable snapshot of all route outputs

The internal project keeps its own operational guide: voynich-codex-project/README.md.

from voynichkit import read_border_matrix, read_exact_forms, FORMS, BORDER

rows = read_border_matrix()                 # full corpus (5,385 loci, 41,005 tokens)
exact = [r for r in rows if r["token"] in FORMS]

print(FORMS)   # ['okal', 'okar', 'okol', ...]
print(BORDER)  # frozenset({'ar', 'al', 'or', 'ol'})

The package exposes the readers read_border_matrix, read_exact_forms, read_zl3b_raw; the constants FORMS, BORDER, OPERATORS, CLOSURES; and the types Token, Locus, MatrixCandidate.

This repository keeps a strict separation between evidence, inference, and hypothesis:

• Do not claim decipherment, translation, or semantic value without independent falsification and source-backed evidence.
• Keep routes, outputs, and next steps documented in docs/summaries/rotas_de_pesquisa.md.
• Update project_state.json whenever a route changes the real state of the investigation.
• Treat visual tooling as support for human review, never as automatic confirmation — never auto-fill human annotation fields from OCR-like inference.

See CONTRIBUTING.md for the full contribution guidelines.

Original code and documentation in this repository are released under the MIT License (see LICENSE). Third-party texts, transcriptions, images, and datasets remain subject to the terms of their original sources — see NOTICE.md and docs/foundations/fontes_e_citacoes.md.

Images downloaded into voynich-codex-project/images/raw/ are git-ignored by default; recreate them locally with the download scripts.