seancollins/voynich-eva-stats

Voynich EVA Corpus Statistics (baseline)

Reproducible descriptive statistics for the Voynich Manuscript in the EVA transcription (Takahashi, IVTFF LSI 2a). A baseline statistics release — a clean, deterministic substrate for hypothesis testing, not a novel finding.

Two streams (this matters)

The word-space is not a written glyph, so all headline quantities are given for both:

• Full character stream — includes the normalized word-space as a token (alphabet 23). Comparable to older EVA-with-spaces figures.
• Within-word stream — space excluded, bigrams formed only inside words (alphabet 22). The linguistically cleaner view.

Quantity	Full stream	Within-word
Glyph tokens	190,954	160,009
H₁ unigram (bits)	3.875	3.861
H(Xₜ, Xₜ₊₁) joint bigram (bits)	6.010	5.714
H(Xₜ₊₁ | Xₜ) conditional (bits)	2.135	1.853

Notation is explicit: H_joint is the joint entropy of adjacent pairs; H_conditional = H_joint − H₁ is bits-per-next-glyph. The low conditional entropy is a well-documented anomaly that any linguistic, cipher, or generative account of Voynichese must reproduce — it is a constraint on explanations, not itself an explanation of cryptanalytic resistance.

Files / configs

• data/glyph_unigrams.jsonl — all glyphs, is_word_space flagged, count + prob.
• data/glyph_bigrams.jsonl — every observed ordered pair (321 of 529 possible), count + prob, so the joint distribution/entropy is reconstructible.
• data/entropy.json — the full/within-word summary above (metadata).

Loads with load_dataset("seancollins/voynich-eva-stats", "glyph_bigrams").

Method

Plug-in Shannon entropy in bits. Miller–Madow bias correction is negligible here (N≈191k, K=23), so it is not separately reported; it matters at higher-order n-grams and small folio-level subsets, which this baseline does not yet cover.

Limitations & what would make it frontier

This is descriptive baseline statistics over one transcription — not a new result. It does not yet include: folio/line/section/Currier-A-B/scribal labels; transcription-uncertainty propagation across Takahashi vs Zandbergen; matched-null generator tests; held-out prediction; or n-gram/compression model comparison. Those — turning the baseline into a transcription-aware, hypothesis-discriminating benchmark — are the paper-worthy next step.

Source: EVA (Takahashi) via the IVTFF LSI file, a community scholarly resource for the public-domain manuscript. Derived statistics released CC-BY-4.0. Companion substrate: SMC17/zsym.