Recovering Voynich Manuscript Bifolium Conjugacy from Textual Similarity

This study tests whether textual similarity can identify documented conjoint partners among competing folio surfaces within the same Voynich Manuscript quire. Using the original Zandbergen–Landini ZL3b IVTFF transcription and published codicological pair labels, 27 bifolia generate 54 oriented retrieval queries across seven quires. Raw EVA token-unigram TF–IDF retrieves the correct partner first in 33/54 cases (61.1%), against a 26.0% average under 20,000 within-quire matching-preserving permutations (one-sided Monte Carlo p=0.00005). A reconstructed 20-feature Controlled representation retrieves 22/54 (40.7%); a leave-one-quire-out blend does not improve on raw EVA. The study builds on prior Layfield–Davis work on conjoint similarity. It does not reconstruct original reading order or independently validate decipherment. This archive includes original input, scripts, archived features, provenance, limitations, and checksums. The original Controlled feature extraction code is unavailable.