Beyond Believability: Evidence-Linked Context Representations for Behaviourally Grounded Social Agents

Social agents can sound human without reconstructing the humans they represent. We test whether an evidence-linked relational-behaviour graph, alone or combined with memory and reflection, improves grounded reconstruction over a compact persona or memory alone. From approximately 230 minutes of edited couples-therapy material, we build representations from earlier episodes and evaluate 400 condition-blind predictions for 5 participants on later held-out situations. Three architectures lead for different participants in this small sample; the hybrid never ranks first, and mechanistic fidelity diverges from observable fidelity. Graphs support provenance and explicit conditional pathways, but errors increase when meaning is carried by surrounding partners or non-verbal behaviour omitted from transcripts. These preliminary findings point toward evidence-calibrated combinations of memory, structured mechanisms and multimodal evidence for behaviourally grounded, psychologically plausible agents. The relevant question is not which persona is most convincing, but which evidence and representation are adequate for the inference being made.