From Parametric Determinism to Emergent Fusion: Data-Curated Style Control in Connectionist Architecture

Abstract This paper interrogates stylistic control in connectionist architectural design through a falsifiable, data-curated design experiment. The ‘cocktail hypothesis,’ which claims that style fusion can be deterministically steered by changing training-data proportions, was tested by training eleven Pix2Pix models on incremental mixes of cultural-facility and recycling-facility façade datasets. Outputs were evaluated through visual review and an independent classifier. At typological extremes, the models reproduced the expected styles consistently; in the mid-range, however, outputs collapsed into a blurred and unpredictable field, falsifying the assumption of linear control. The analysis treats this ‘indeterminate middle’ not as an error but as a signature of generative process—an emergent style fusion aesthetic arising from subsymbolic features rather than additive collage. Ultimately, these findings shift architectural authorship from deterministic control to the curation of emergence, a practice defined by the deliberate selection, proportioning, and weighting of data.