A quantitative, semantics-neutral benchmark for the Pisces subsystem of the Voynich Manuscript. Includes clustering hierarchies and spatial dispersion statistics to test generative text models. Abstract: Decipherment claims for the Voynich Manuscript frequently fail due to a lack of reproducible structural benchmarks. This paper presents a semantics-neutral analysis of the 30 labels on folio f70v2 (Pisces), establishing a multi-layered profile of the text’s behavior. Using hierarchical agglomerative clustering (HAC) and Levenshtein distance, we identify tight morphological paradigms—notably an otal- family and an ok/ot- + -dy/oly family—alongside stable structural outliers. We extend this analysis with positional dispersion metrics, revealing that while some paradigms exhibit regional arc-biases, others (the otal- group) are over-dispersed globally around the wheel. These patterns, contrasted with the rapid clustering collapse and low entropy of the qo- dominant inner rings, suggest a non-local, rule-based generative mechanism. We provide a full dataset and pipeline as a falsifiable standard; any viable model for the manuscript must not only match global statistics but also reproduce the specific clustering hierarchies and spatial asymmetries identified in this subsystem.