Voynichese Has Natural-Language-Range Character Entropy but Anomalous Order-Scaling: A Corrected Statistical Characterization

⚠️ Correction (v2, July 2026): this version withdraws the abbreviated-Latin decipherment claim of the original. The χ²=41.56 result tested a uniform-null strawman (≈358 on the real counts) and is no evidence for Latin; the "five-case system" was post-hoc relabeling that excluded the single most common word ending (37.9%); the "94.1% accuracy" used unfalsifiable scoring; the transcription was v101, not EVA; and a plant-label decode test failed. Measured correctly, Voynichese has natural-language-range character entropy but anomalously flat within-word order-scaling, and does not support a specific Latin reading. I no longer hold that reading. I re-analyze the Voynich Manuscript text (Yale Beinecke MS 408) in the v101 (Glen Claston) transliteration — ~40,600 word tokens after cleaning — using held-out, smoothed character n-gram models (interpolated absolute discounting) computed by an identical pipeline against real Latin and Italian controls, with confidence intervals from repeated splits. Findings: (1) Voynichese's character entropy is within the natural-language range, and at equal corpus size it is the least predictable of the four (order-3 perplexity ~5.9–6.1 vs Latin 4.7 and Italian 4.9) — "too predictable to be a language" is not supported. (2) Its within-word order-scaling is anomalously flat: predictability barely improves as character context grows (Δ(H₁−H₃) ≈ +0.08), whereas three natural-language controls all gain ~1 bit (Δ ≈ +1.0). (3) About half of that gap is explained by a mundane null the original never ran: vowel-stripping the controls (an abjad/vowelless orthography) collapses their Δ from ~1.0 to ~0.45 — yet a real residual survives, since Voynichese (+0.08) is flatter than even devowelled natural language (+0.45). (4) The effect is strictly within-word: cross-word character information is ≈0 in every corpus including natural language, so these character-level statistics are silent on syntax — and a token-level check finds only slow long-range drift, not sentence-level syntax. (5) The word-final inventory is highly skewed — one ending is 37.9% of all words — dissolving v1's "five-case Latin system" into a selective relabeling — and the affixation is templatic (stem→suffix mutual information 0.29 vs ~0.86 for Latin/Italian), not the paradigm-blocking of real declension. (6) Glyph frequencies are strongly layout-dependent — specific glyphs are enriched 14–22× at line starts and ends (the bias vanishing under a within-line shuffle), so the written line is a structural unit, which running prose cannot produce. A botanical-label decode test from the original materials fails. The text is consistent with a constrained generative system whose strong within-word constraint is partly attributable to vowelless-style orthography; the data do not license a specific natural-language grammar. These results largely replicate established work; the bounded within-word order-scaling anomaly and its abjad null are the new contributions.