Word-Boundary Sensitivity of Voynich Zipf Null-Model Tests

Whether a generative null model for the Voynich Manuscript's Zipf-like rank–frequency slope "reproduces" the observed value depends on which claim the null is actually being asked to support: that the slope survives when the model must place word boundaries itself (G, boundary-generated), or only that it survives when the model is told where the observed boundaries already are (I, boundary-conditioned). We preregistered this as a confirmatory test (OSF, DOI 10.17605/OSF.IO/DE3CN) on three fitted character-level Markov null orders (N0–N2, order 0–2, Laplace-smoothed), sampled by exact forward-filtering/backward-sampling under both regimes from one and the same fitted model, across the pre-specified 72/12/36-effective-cell grid (ZL3b, IT2a, GC2a; IT2a's uncertain-space axis is degenerate) with R = 30 realisations per null × regime × cell. The primary estimand, F_N — the proportion of paired-evaluable effective cells where the I- and G-verdicts differ — is order-dependent and large where it is non-zero: F_N = 0.72 at N0 and 0.97 at N1 (both H1-supported, F_N^stable in close agreement), collapsing to F_N = 0.00 at N2 (insensitive) with no evaluability exclusions at any order (all 360 null × cell pairs paired-evaluable). At N0 and N1 the boundary-generated regime systematically fails to reproduce the observed slope (median |D| ≈ 0.29–0.38, τ = 0.20) while the boundary-conditioned regime typically passes (median |D| ≈ 0.05–0.10); at N2 both regimes converge to small D and pass almost everywhere. Realisation-level R² stays high and near-identical across regimes and orders (median 0.96–0.98), so this pattern is not a fit-quality artefact of the amended validity rule. The study-level conclusion is H1: the Zipf-null verdict for Voynichese is boundary-regime-dependent at low model order and becomes regime-independent as the fitted model is given enough context to place boundaries on its own. No decipherment, language identification, or authorship/origin claim is made or supported. Companion software/data deposit (code, engineering checks, confirmatory run output, aggregation): Zenodo, DOI 10.5281/zenodo.22866913. Preregistration: OSF, DOI 10.17605/OSF.IO/DE3CN. Submitted to Digital Medievalist.