sdaffdsaffdsafsd/voynich-encipherment-structure-public-data-20260418

Voynich Encipherment Structure Public Data

This package is the public-data supplement for Reading the Mechanism, Not the Message: Studying Linguistic Structure and Mechanism in the Voynich Manuscript.

It contains generated research artifacts that are useful to other researchers:

• BERT-family function-shell, historical Italian/Romance, lexical-class, and generated dataset outputs.
• 15-page test-cluster POS/tagging, Italian-grounded token ledgers, and BERToldo masked-language-model candidate lattices.
• Mamba token-syntax event streams, training metrics, confusion matrices, and ablation courts.
• Compact-unit, entropy-patch, token-compressor, finite-state transducer, surface compiler, exact replay, serializer, and no-plaintext court outputs.
• The current LaTeX paper bundle, source manifests, synthetic EM/transformer controls, and corpus/count audits.
• A supplementary evidence crosswalk from manuscript citation keys to public paths/checksums, plus an archive-risk manifest for provenance that is outside this paper supplement.
• A small manifest validator in files/paper/validate_public_package.py.

Public Locations

• Hugging Face dataset: sdaffdsaffdsafsd/voynich-encipherment-structure-public-data-20260418
• Hugging Face URL: https://huggingface.co/datasets/sdaffdsaffdsafsd/voynich-encipherment-structure-public-data-20260418
• GitHub branch: https://github.com/createdaccountbecauseIwantgithubcopilot/Voynich/tree/voynich-paper-public-data-20260418

Claim Boundary

This is a mechanism/structure supplement. It does not authorize public plaintext, named-language proof, or a candidate translation. The included tables intentionally preserve no-plaintext and failed-promotion records. Candidate-language labels in lattice/probe files are diagnostic payloads and must not be read as translations.

Layout

• files/ groups artifacts by evidence family under neutral public paths.
• manifest_public_data.tsv lists every included file with byte count and SHA-256 checksum. It uses evidence-family labels and source fingerprints instead of private run-directory paths.
• large_model_artifacts_not_included.tsv lists BERT-family and neural checkpoint/cache binaries that were intentionally not included in this GitHub-ready package. The generated outputs and model/source manifests are included, and excluded artifacts are identified by source fingerprint rather than private path.
• External-corpus plaintext snippets and target-token payloads in generated language-prior datasets are redacted from this public package. The synthetic cipher-side features, metrics, labels, and checksums remain.
• public_data_summary.json gives a machine-readable package summary.
• files/paper/supplementary_evidence_crosswalk.tsv maps manuscript bibliography/supplement keys to public paths and SHA-256 values.
• files/paper/archive_risk_manifest_20260419.tsv lists local provenance that should be archived separately before deleting the private workspace.

Reproduction Notes

The package excludes scratch directories and manuscript images/PDF files. The generated paper PDF is included when present beside the LaTeX source; this is not a manuscript-image exception. The paper's own reproducibility protocol is available at:

files/paper/reproducibility_protocol.md

The underlying manuscript transcription used by the local audit is the cached Ched-ai/voynich-eva dataset snapshot recorded in the paper bundle.

This package is not a complete provenance archive for the broader private workspace. In particular, late frontier bundles, visual/page-routing assets, and large model/corpus binaries need separate archival if exact future reruns are required.