# Voynich Harmonic-Topological Analysis — Reproducibility Repository **Paper:** *Graph-Hodge Analysis of the Voynich Manuscript as Generator-Class Discrimination* **Author:** Derek Earnhart (OriginStoryModel) **Date:** April 2026 --- ## Reproducibility Scope — Read This First This repository has **two reproducibility tiers**. It is critical to understand which tier you are using: ### Tier A — Processed CSVs only (fully self-contained) > **The inferential BEH table, null-control verification, and all reported statistics can be reproduced from the processed CSVs alone. The raw IVTFF file is NOT required for this tier.** This is the primary reproducibility path. All seven `data/processed/*.csv` files and `data/processed/summary.json` are included in the repository. Running the two commands below is sufficient to regenerate every number in the paper: ```bash python src/entropy_tests.py --processed_dir data/processed --results_dir results \ --permutations 10000 --seed 42 python src/null_controls.py --processed_dir data/processed --results_dir results ``` ### Tier B — Full raw IVTFF re-extraction (requires raw file download) > **If you want to re-run the IVTFF parsing and feature extraction from scratch, you must download `LSI_ivtff_0d.txt` separately.** The raw file `data/raw/LSI_ivtff_0d.txt` is NOT included in this repository (it is a third-party file maintained by the open Voynich transcription community). See [Downloading the Raw IVTFF File](#downloading-the-raw-ivtff-file) below. Once the raw file is in `data/raw/`, the full pipeline is: ```bash python src/parse_ivtff.py --input data/raw/LSI_ivtff_0d.txt --transcriber H python src/clean_eva.py --input data/processed/tokens_clean.csv \ --output data/processed/tokens_normalized.csv python src/transition_graph.py --tokens_csv data/processed/tokens_clean.csv \ --output_dir results/ python src/hodge_core.py --tokens_csv data/processed/tokens_clean.csv \ --output_dir results/ python src/entropy_tests.py --processed_dir data/processed --results_dir results \ --permutations 10000 --seed 42 python src/null_controls.py --processed_dir data/processed --results_dir results ``` --- ## Downloading the Raw IVTFF File The Landini-Stolfi IVTFF EVA archive is maintained by the Voynich transcription community. **Primary source:** https://www.voynich.nu/data/beta/LSI_ivtff_0d.txt **Expected SHA256 checksum:** ``` 3f3f2af18cde10efe75c582f49b07b651c3397022fcbfa5854fecc424c121afa LSI_ivtff_0d.txt ``` Verify before use: ```bash sha256sum data/raw/LSI_ivtff_0d.txt ``` The checksum must match exactly. If it does not, the file version may differ from the one used to generate the processed CSVs, and results may differ slightly. --- ## Repository Layout ``` voynich_repo/ ├── README.md ← this file ├── LICENSE ← MIT license for repository code ├── requirements.txt ← pip dependencies │ ├── data/ │ ├── raw/ │ │ ├── SHA256SUMS.txt ← checksums for all files │ │ └── LSI_ivtff_0d.txt ← NOT included; download separately │ │ │ └── processed/ ← INCLUDED; sufficient for Tier A │ ├── aggregate_features.csv ← full-manuscript and stratum-level features │ ├── page_features.csv ← per-page entropy and graph features (225 pages) │ ├── paragraph_unit_features.csv← per-paragraph-unit features │ ├── matched_controls.csv ← 20-trial frequency-shuffle + uniform-vocab controls │ ├── matched_controls_summary.csv │ ├── top_tokens.csv ← most frequent EVA word tokens │ └── summary.json ← machine-readable extraction summary │ ├── src/ │ ├── parse_ivtff.py ← IVTFF parser (Tier B only) │ ├── clean_eva.py ← EVA token cleaning utilities │ ├── transition_graph.py ← token-transition graph construction │ ├── hodge_core.py ← Hodge 1-Laplacian decomposition │ ├── entropy_tests.py ← BEH inferential statistics (primary script) │ └── null_controls.py ← null-control verification │ ├── results/ │ ├── beh_inferential.csv ← completed inferential BEH table │ └── beh_tables_publication.md ← Markdown + LaTeX ready for manuscript │ └── preregistration/ ├── analysis_plan.md ← pre-specified hypotheses and thresholds └── thresholds.yaml ← machine-readable pass/fail criteria ``` --- ## Quick Start (Tier A — 2 commands) ```bash # 1. Install dependencies pip install -r requirements.txt # 2. Regenerate inferential table python src/entropy_tests.py \ --processed_dir data/processed \ --results_dir results \ --permutations 10000 \ --seed 42 # 3. Verify null controls python src/null_controls.py \ --processed_dir data/processed \ --results_dir results ``` Expected output of Step 2 (also regenerates results/beh_tables_publication.md): ``` Computing: Currier A vs B / page n=114/82 Computing: Currier A vs B / paragraph n=141/111 Computing: Hand 1 vs 2 / page n=86/45 Computing: Hand 1 vs 2 / paragraph n=92/64 Two-sided BEH supported (all q < 0.05): True Original one-sided direction (A/1 > B/2) failed: True ``` --- ## Key Results (pre-computed in `results/beh_inferential.csv`) | Comparison | Level | n₁/n₂ | ΔH | 95% CI | p (2s) | q (BH) | g | δ | |---|---|---|---|---|---|---|---|---| | Currier A vs B | page | 114/82 | −0.943 | [−1.123, −0.762] | < 0.001 | < 0.001 | −1.544 | −0.707 | | Currier A vs B | paragraph | 141/111 | −0.695 | [−0.901, −0.484] | < 0.001 | < 0.001 | −0.870 | −0.457 | | Hand 1 vs 2 | page | 86/45 | −0.750 | [−0.908, −0.595] | < 0.001 | < 0.001 | −1.810 | −0.780 | | Hand 1 vs 2 | paragraph | 92/64 | −0.351 | [−0.557, −0.133] | < 0.001 | < 0.001 | −0.578 | −0.381 | **Two-sided BEH: SUPPORTED** (all BH-adjusted q < 0.001) **Original one-sided BEH (A/1 > B/2): REFUTED** (all observed ΔH < 0) --- ## Checksum Verification To verify all included files: ```bash # From repo root — verifies data/processed and results files only # (data/raw/LSI_ivtff_0d.txt line will fail if raw file not present; that is expected) sha256sum -c data/raw/SHA256SUMS.txt 2>/dev/null | grep -v "No such file" ``` All included files should report `OK`. The raw IVTFF line will report `No such file or directory` if not downloaded — this is expected for Tier A users. --- ## Randomness and Reproducibility All repository stochastic operations use deterministic NumPy generators seeded with 42. The primary inferential/statistical scripts use `numpy.random.default_rng(42)`, and there are no calls to the legacy `numpy.random.seed()` interface. Output files include: - Random seed: 42 - Permutations: 10,000 - Bootstrap resamples: 10,000 Results should be bit-for-bit identical across runs on any platform running NumPy ≥ 1.24 with the same seed. --- ## Manuscript Update Guide After regenerating `results/beh_inferential.csv`: 1. **Must:** Replace Table 7 (Section 9.4) with the LaTeX table in `results/beh_tables_publication.md`. 2. **Must:** Add footnote disclosing paragraph-level discrepancy: *"Paragraph-level Currier B mean = 6.407 (computed) vs. 6.453 (earlier draft); Hand 2 paragraph mean = 6.077 (computed) vs. 6.115 (earlier draft). Counts differ by 1–2 units due to paragraph-boundary definitions. Directional finding unchanged."* 3. **Must:** Update Reproducibility Statement with public repo URL and SHA256SUMS reference. 4. **Recommended:** Add to Abstract: two-sided BEH confirmed with q < 0.001 and effect sizes g = −0.578 to −1.810. 5. **Recommended:** Add to Section 9.2: biological section has the highest successor entropy (3.734) of any section, supporting structural heterogeneity across domains. --- ## Dependencies ``` numpy>=1.24 pandas>=2.0 scipy>=1.11 ``` `scipy` is required only for `hodge_core.py` (sparse linear algebra). The primary inferential script `entropy_tests.py` uses only `numpy` and `pandas`. --- ## License Repository code is released under the MIT License; see `LICENSE`. The IVTFF transcription data is maintained by the Voynich transcription community and is not covered by this license. --- ## Citation Earnhart, D. (2026). *Graph-Hodge Analysis of the Voynich Manuscript as Generator-Class Discrimination: Lossless Glyph Encodings, Substitution-Invariant Transition Flows, and Falsifiable Structural Tests.* [Preprint]. Repository: [INSERT OSF/ZENODO/GITHUB URL BEFORE SUBMISSION]