Shannon’s decoder recovers the original signal. Its objective function is fidelity. Modern decoders — transformers — learn their objective function from training data, producing outputs shaped by that learned objective rather than by fidelity to a ground truth. The relational decoder is a predictive processing system — a generative model minimizing prediction error — whose error-minimization is constrained by social pain. Social pain recruits overlapping neural regions implicated in physical pain processing — though later work indicates partly distinct representations at the multivariate pattern level (Woo et al., 2014) — and the system appears to have been selected to minimize it. This is an adaptive interpretation consistent with the exaptation framework (Gould and Vrba, 1982) rather than a directly tested evolutionary history. The constraint produces systematic, directional departures from accurate reconstruction that predictive processing alone does not predict. Given a flow signature — the information actually arriving from a source — the relational decoder does not ask what the source is transmitting. It asks what reconstruction minimizes pain. Available flow is attended to selectively, amplified where it supports low-pain reconstruction, suppressed where it does not, and supplemented by internally generated signal where it is absent. The result is felt relationship — an output of the decoder, shaped by its objective function, attributed by the decoder to the source. This account unifies grief, parasocial attachment, and religious experience within a single framework. It explains why felt relationship persists in the absence of signal, why disconfirmation resistance scales with prior depth, and why the model’s architecture extends to AI companion systems in ways that existing frameworks do not anticipate. It generates testable predictions that existing frameworks — attachment theory, predictive processing, anthropomorphism research — do not individually produce. The relationship is not in the object. It is in the decoder.
Lloyd Taylor (Tue,) studied this question.