The Voynich Manuscript (Beinecke MS 408) has carried, for more than a century, a parallel literature claiming the surface text is the output of a constructed cipher, either a genuine cipher of pre-1500 vintage or a later forgery applying a known cipher method. Every prior cipher trial in the public literature has used a single acceptance criterion: lexical recovery against a natural language dictionary. Under that criterion, every cipher trial returns NULL, and the result is read as "not Latin" or "not German" rather than as a structural constraint on the surface. This paper changes the acceptance criterion. We define a substrate operator grammar scorer over the chip class lexicon attested across the manuscript, then run the catalog of known cipher methods from antiquity through 2026, including representative modern computationally secure primitives (DES, AES, ChaCha20) and a representative post-quantum sample (Kyber, NTRU, McEliece, BIKE, HQC, plus structural simulation rows for SPHINCS+; Dilithium and other signature schemes are excluded from the cipher catalog as categorically non-encryption), across thirteen folios drawn from six structural sections (cosmographic, herbal, balneological, pharmaceutical, cosmological, zodiac, rosettes foldout). Across 1014 v3 (cipher, key) trials at catalog scope, zero combinations produce output that scores higher than the unciphered manuscript under the substrate grammar scorer. A separate generator-side test of the Cardano grille hypothesis (Rugg 2003) returns zero of six tested grille configurations achieving distribution match against any real folio across six distribution statistics, with a structural argument (Section 7.4) that any grille reading multi-chip syllables is forced to over-produce token length by construction. The structural argument for modern strong ciphers is reframed: the substrate scorer is not a cryptographic distinguisher against uniform random; it measures alphabet distribution match against the manuscript's empirically non-uniform opener class distribution, and any cipher whose output is uniform over the chip alphabet falls to ~10/22 mean opener fit by construction. The empirical 25-45 percentage point gap between unciphered manuscript scores (0.74-0.98) and the null floor of modern cipher output (0.32-0.58) closes the entire class by this alphabet distribution mismatch (Observed). The constructed cipher class of forgery hypotheses is closed at this evidence level across the tested folios. A v4 extension (Section 10) re-grounds the language scope in substrate evidence of the 1450s Hartlieb tradition language stack (Hebrew, Arabic, Greek register switching, Mittelhochdeutsch, Latin medical / Cappelli, Mandeville tradition exotic alphabets, namenmantik, sign legend, alchemical sigla) and adds a second scorer dimension for shorthand fit. The v4 extension is largely co-witness at the alphabet-size-parametrized substitution layer (per the Section 10.1 scope note: medieval Hebrew / Arabic / Greek letter-to-Voynich-chip assignment is not substrate-attested at the per-letter level), and confirms that even when alphabet sizes are tuned to the Hartlieb-tradition substrate-grounded scope (22 / 28 / 24), the closure holds at substrate-grounded alphabet sizes (0 grammar hits across the v4 catalog scope). A separate observation about the al-Manazil mod-28 block (Block B) is reported under explicit scope: a position cyclic substitution is mathematically forced to score volvelle traversal strength 1.0 on any input of sufficient length, so the Block B strength=1.0 result is a scorer cipher tautology and is reported as a methodology landmark rather than as substrate positive evidence. A separate, content-dependent candidate signal at corpus-complete scope is the manuscript's native unciphered mod-28 preference, which (at 224 folios tested) holds on 38.4% of eligible folios with strong section heterogeneity (astronomical 62.5%, pharmaceutical 56.2%, zodiac 50.0%, cosmological 20.0%, herbal 35.2%), a 42.5 percentage point section spread that meets the dramatic divergence threshold; this is reported as a candidate signal pending a per-folio shuffled-content null and section-stratified covariate-controlled comparison. The result does not address authenticity (handled by the 2009 Arizona AMS C14 dating and McCrone ink analysis), does not claim the manuscript encodes natural language, and does not exclude information-theoretic one-time-pad use with a uniform pad.
Honeycutt, Edwin Marshall, III (Thu,) studied this question.