Does page similarity reveal quire structure? A controlled test of the reordering procedure proposed for the Voynich manuscript

Layfield and Davis (2026b) proposed that the Voynich manuscript (New Haven, Beinecke Rare Book and Manuscript Library, MS 408) was conceived as a series of single-bifolium gatherings (singulions) and later bound out of order. Their codicological and palaeographical case is supported by latent semantic analysis (LSA) of page similarity, used to compare conjoint, facing and confoliate pages and, for two quires, to find the order of bifolia that maximises similarity. Their similarity criterion was checked on two codices of known structure, but the reordering step had not been tested on any text of known order. Under preregistered designs, revised after adversarial review, we reconstructed the procedure and applied it to two manuscripts of known nested structure, the Auchinleck manuscript and AM 619 4to. Where text runs continuously across pages, the procedure recovers the true nested arrangement of a quire (mean normalised rank 0.035 over five quires, p = 4 × 10⁻⁵), robustly to every preregistered sensitivity. TF-IDF and Jaccard baselines do as well, so the signal is carried by vocabulary shared between neighbouring pages. The question-1 criterion identifies a third known nested codex and returns no false singulions. The improvement from reordering as singulions, the quantity that selects the proposed orders, is not unusual under nulls built from pages without continuity: the published +12.1% and +20% lie between the 9th and 14th percentiles at their own quire sizes, whereas in real nested quires with continuity the true order beat the best singulion order on average. By itself, then, the improvement does not single out the proposed orders; its distribution under misbound singulions with continuity was not built. A second test asked whether, in nested codices copied by one hand, conjoint leaves resemble each other beyond reading distance. Across twelve Castilian volumes the estimate was Δ = +0.048 (90% interval −0.087 to +0.165), against mean values of +0.11 and +0.54 simulated before measurement for moderate and large bifolium effects; the preregistered rule had no attainable margin, so this test is without conclusion. Nothing was measured on the Voynich. Preprint of an article submitted to Digital Medievalist on 27 October 2026. Code, preregistrations and results: https://doi.org/10.5281/zenodo.23687541.