Voynich decode — strict visual observers (weights)
Private backup of the computer-vision "observer" models used in the Voynich (Beinecke MS 408) structural-factorization project. These models read glyphs from the Yale scans to confirm disputed cells for "strict coverage". They are not a decipherment: the result itself (factorization + a sufficient statistical generator) is produced by plain count-statistics, not by these nets.
Honest scope. The project reconstructs and reproduces the writing mechanism, not a message. No plaintext / translation is claimed; whether the text is a hoax or a very verbose cipher is undecidable from the text alone.
What is here
Mirror of MASTER/models/ (weights + registry + cards + gate summaries).
Accepted (passed frozen OOF / 999 matched-null / independent audit):
| model | role | strict additions |
|---|---|---|
leaf_clean_highres_shadow_gate_v1 |
Yale high-res dual-observer surface | 272 |
strict_v2_broad_shadow_gate_v1 |
frozen highres observer, broad queue | 801 |
strict_v3_consensus_gate_v1 |
consensus gate (≥2 observers) | 59 |
strict_v4_direct_factor_gate_v1 |
direct visual factor | 679 |
strict_v5_short_factor_gate_v1 |
short exact-factor specialist | 304 |
strict_v6_continued_direct_gate_v1 |
continued direct, calibrated threshold | 74 |
strict_v15_image_candidate_ranker_v1 |
image-only candidate ranker | 8 |
Plus engineering_baselines/ (warm-start / audit, not strict) and
quarantine/ (failed content gate). Architectures: CRNN/CTC,
patch-transformer + semi-Markov, metric/embedding, gradient-boost/RF rankers.
Each accepted model = 15 physical-leaf OOF checkpoints (5×3 fold/seed).
Intended use & limitations
- Use: re-running glyph/cell inference on Beinecke MS 408 scans to reproduce the strict factorization audit.
- Not for: any "reading"/translation claim; general handwriting OCR; other manuscripts (trained on one manuscript's hand).
Training data & provenance
- Images: Beinecke MS 408, Yale University Library — public domain (Yale asserts no copyright; images via IIIF). Courtesy credit: Beinecke Rare Book & Manuscript Library, Yale University.
- Labels/transcription: Zandbergen–Landini EVA interlinear (voynich.nu), with project corrections — © R. Zandbergen; attribute, do not re-host raw.
- Cipher benchmark (ancillary, failed solver attack): DECODE/DECRYPT database — CC BY 4.0 (Megyesi et al.).
- EVA font (display only): © G. Landini 1998, free/non-commercial.
The weights themselves are the authors' trained parameters, released CC-BY-4.0; third-party data above is not redistributed here — fetch from source.
Use
pip install -U huggingface_hub
hf auth login
python -c "from huggingface_hub import snapshot_download; \
snapshot_download('LeonidMolchanov1987/voynich-decode-models', repo_type='model', local_dir='MASTER/models')"
Integrity: compare SHA-256 against the bundle's WEIGHTS_MANIFEST.tsv.
Proof / methodology
Evidence (null-gates, generator, reproducible runner prove_bundle.py) lives in
the project's code/proof bundle (GitHub / Zenodo DOI — link to be added).
Author: Молчанов Леонид Олегович.