Relational Bayesian Inference for Musical Pleasure: An Experimental Protocol

Can a listener’s response to a question about familiarity or expectation improve the prediction of their subsequent musical pleasure? This protocol proposes a test within Musical Intelligence (MI), using a fixed audio representation and a Bayesian observation model with a persistent listener state. History, direct pleasure reports and auxiliary responses enter separate comparisons, allowing the contribution of relations between tasks to be estimated on the same listening trajectory. Predictions are committed before each listening block and evaluated on later reports. A conditional Gaussian example establishes when an auxiliary observation can reduce predictive uncertainty, while a wrong-sign example shows how an incorrect relation can worsen prediction. The experimental design couples this comparison to a separate, controlled search for associations beyond the dimensions already labelled as pleasure-related. It specifies stimulus provenance, temporal alignment, model controls, held-out source families and simulation-based planning. The personal inference layer and experiment are proposed extensions; the figures show source-audited architecture, analytic calculations and synthetic examples, not participant results.