Corrigendum and Concord Control: Corrections to the v15 Companion (Voynich Manuscript)

Version 16 is a corrigendum-and-control note following an external cross-model red-team of the consolidated findings. It corrects one claim in the v15 companion and adds a new real-language control. The structural paper remains unchanged and under review at Cryptologia (Submission ID 265719422); the frozen substitution table is neither used nor modified. Correction. The v15 companion stated that an unmodified self-citation generator reproduces the manuscript's long-range mutual-information plateau and near-neighbour repetition rate out-of-sample. A multi-seed audit shows that conclusion rested on single-seed output: across eight distinct seeds the generator over-produces the MI plateau (0.044±0.022 vs real 0.013) and near-repeats (0.071 vs 0.048), under-produces word length (4.55 vs 4.99) and type/token ratio (0.121 vs 0.203), and reaches ~54% of the d=1 spike — the real value lies outside the generator's across-seed range on every metric, and no single configuration reproduces the signatures jointly. The "reproduces out-of-sample" claim is withdrawn; the supported statement is that self-citation produces statistics of the right character and order of magnitude but does not robustly reproduce the values. Self-citation is a plausible kind of mechanism for the structure, not a fitted reproduction of it. New control. A Latin prose corpus (Caesar, Cicero; ~57,600 tokens) run through the identical d=1 pipeline shows a coda/ending-dominated spike (onset/coda ratio 0.29) — the signature of suffixal concord — whereas the Voynich spike is onset-dominated (~5.8). The morphological frame reconstructs a high fraction in both languages (Latin 83%, Voynich 72%), so that fraction does not discriminate; only the onset-vs-coda location does. The Voynich's d=1 agreement is therefore unlike the suffixal concord of the Latin/Romance languages proposed for the manuscript (prefixal-agreement languages not excluded). Supporting checks (vocabulary-matched MI; a cipher positive control showing attestation-density scoring detects true and partial solves; space-handling robustness) and tier/framing corrections are included. The central conclusion — convincing procedural structure without demonstrated referential content — is unchanged, and now rests explicitly on real-text analyses rather than on generator reproduction. The structural paper and the v15 companion are both retained in this version.