Trust as Predictive Precision: Reliability and Influence in Representation Alignment

Work on trustworthy machine learning often invokes trust, confidence, and reliability without specifying how these quantities relate. We study a predictive-coding model in which agents exchange noisy messages generated from latent world representations and learn a directed precision parameter by minimizing expected log loss. In this setting, trust is the calibrated precision assigned to another source's messages after accounting for representation mismatch and irreducible source noise. Optimizing this precision yields a closed-form trust kernel. Homophily-like influence, asymmetric teacher-student alignment, and consensus dynamics then follow from the same residual model. The result is a restricted but testable account of epistemic trust in representation alignment: agreement supports trust only when it is mediated by calibrated predictive reliability.