A Relational Context Layer: A Reference Architecture for AI Support of Response-Mediated Work

Large language models are stateless with respect to the user's relational world: each interaction begins without a persistent model of the people the user must work with and through, of the user's own goals and constraints, or of the structural conditions that bound what the user can achieve. This caps the quality of relational assistance, and it is an architectural property rather than a deficit that scale will close. I propose the Relational Context Layer (RCL), a reference architecture for AI systems supporting response-mediated work — professional tasks whose success is mediated by other people's responses, whether or not the work is interpersonal in form. RCL integrates a first-person model of the served human (navigator layer), other actors held as context-conditioned estimates of what they want (relational layer), and the structural conditions bounding the possibility space (contextual layer), maintained over the accumulated record of the relationship. Formally, RCL is expressible as a finitely nested interactive POMDP under an epistemic reward; I state this mapping rather than contest it. Its central bet, which no reduction can absorb, is longitudinal: a system supporting such work predicts worse — most sharply on the actions whose meaning is history-constituted — if it severs want-inference from that record. Claims carry explicit maturity tags, and the validation programme specifies corpus, metrics, and decision rules in advance. An early prototype, ReactAI, in which I have a commercial interest, is noted as provenance only.