Glossary

Plain-language definitions for the two audiences this project serves — Voynich/MS 408 researchers who may not be coders, and technical users who may not know the manuscript. Where a term names a computed quantity, the “computed in” note points at the code.

The manuscript

• Voynich Manuscript / Beinecke MS 408 — an illustrated codex in an undeciphered script, held at Yale’s Beinecke Library (shelfmark MS 408); carbon-dated to the early 15th century.
• Folio — a numbered manuscript leaf (e.g. f73r = folio 73, recto/front). Sections are conventionally labelled herbal, astronomical, balneological (“bathing”), pharmaceutical, etc.
• Hand / scribe — a distinct handwriting. Multiple scribal hands have been identified (Fagin Davis); text analyses stratify by hand.
• Currier A / Currier B — two statistically distinct “languages” or systems within the text (Prescott Currier). Whether they are dialects, registers, or vocabularies is open; all our text analyses stratify by A/B.

Transliteration

• Transliteration — a rendering of the script’s glyphs into ASCII so it can be computed on. The choice of scheme can change statistics, so we treat it as a variable.
• EVA — Extensible (originally European) Voynich Alphabet; the primary transliteration we use (Zandbergen–Landini, “ZL”). v101 is an alternative scheme used as a sensitivity check.
• IVTFF — Intermediate Voynich Transliteration File Format; the file format the ZL transliteration ships in (parsed in src/ms408/ivtff.py).
• Respacing — re-drawing word boundaries. It changes token statistics dramatically and is a known confound (a cipher can respace its plaintext before enciphering) — see the E29 note.

The discriminator axes (what evaluate() reports)

• h2 — conditional (second-order) character entropy: how predictable the next character is given the previous one. The manuscript’s is anomalously low (~2.08) vs natural language (~3.1–3.9). A “hard” axis. Computed in textstats.lb_entropies (Lindemann–Bowern).
• ΔI (dI) / MZ information — Montemurro–Zanette word-order information: how much information the order of words carries at a given scale. Confounded: it collapses under homophony or respacing alone, so it is a homophony/type-token detector, not a clean word-order axis — never read an in-band ΔI as “intact word order.” Computed in src/ms408/mz.py.
• ed1 / ED1 — edit-distance-1 morphology: the share of word types connected in a network where an edge joins words one character-edit apart. The manuscript has a dense such network (~0.75–0.80). A “hard” axis. Computed in studies/morphology.py.
• Zipf slope — the slope of the rank–frequency law (log rank vs log frequency). A “hard” axis. Computed in textstats.zipf_slope.
• TTR (type–token ratio) — vocabulary richness = distinct words ÷ total words. Advisory: intrinsically sensitive to the token count (fewer tokens → higher TTR), so it is not banded and not counted; only compare at a similar token budget.
• fc_z / wc_z (mid-level syntax) — null-corrected z-scores for the function/content collocation gap (fc) and adjacent word-class structure (wc). “local” uses a within-block null (deconfounded of topic drift); “global” uses a full order-shuffle null (confounded by topic drift). Both are soft: the manuscript’s own confidence interval crosses zero. Computed in signature.py (from e13b/e14 primitives).
• block scale — a coarse partition of the text (hundreds of words), where topic/section structure lives, as opposed to the token/adjacency (word-to-word) scale.
• null-corrected — a statistic expressed relative to a matched random baseline (e.g. an order-shuffle of the same words), so nuisance factors cancel and corpora are comparable.

The method

• Harness — matched controls (real language, ciphers, meaningless generators) that a method must separate on synthetic data before it is trusted on the manuscript.
• Firewall — the rule that every reported number comes from deterministic, versioned code writing machine-readable results; nothing estimated or recalled.
• Refutation pass — an independent, clean-context skeptic tasked to attack a strong claim before it stands; briefs archived in docs/refutations/. See docs/METHODOLOGY.md.
• Evidence grade (A–D) — the strength label every claim carries; A is replicated/robust, D is speculative. An ungraded claim is treated as D.
• Necessary, not sufficient — matching the manuscript’s bands means a hypothesis is not excluded, never that it is the mechanism. The evaluator is built to withhold, not confirm.