Many voices in philosophy, sociology and heterodox economics—from Continental philosophy, to more recent critiques of complexity science—have long challenged reductionism and scientism in the social sciences. In particular, psychology has faced criticisms for "naturalizing" contingent social conditions, cloaking its socially constructed nature under a façade of objectivity. At a methodological level, scholars such as Denny Borsboom, Joel Michell, and Sara Dellantonio have highlighted the arbitrariness and rigidity of psychological constructs, as well as the circularity and reductionism embedded in the item pre-structuring of psychometrics. These critiques signal an urgent need for a paradigm shift in psychometrics and sociological survey analysis.

We propose novel statistical tools and methods that could guide researchers in such a paradigm change. We aim, in particular, to tackle three objectives: (1) assessing the descriptive and explanatory power of different item sets—enabling the construction of optimal, highly informative reduced questionnaires or fusions of existing ones; (2) detecting structural artifacts introduced by specific item formulations, such as artificial clusters in response data induced by item redundancy; and (3) disentangling trivial covariances caused by semantic overlaps in questions from non-trivial covariances that reflect genuine latent traits. We argue that these three intertwined objectives can be addressed under a single, principled approach. Grounded in information theory, statistical inference, and statistical physics, our approach moves beyond classic criteria based solely on internal consistency (Cronbach, 1951), optimizing the trade-off between randomness and redundancy of sets of items, and ideally filtering out both noise and redundant items.