BEYOND SYCOPHANCY: MODELING RELATIONAL INTEGRITY IN CONVERSATIONAL AI

The system operates as a constrained generation pipeline. User input flows through affect analysis, which detects emotional state and maps sentiment to a five-dimensional relational vector: hurt, trust, hope, frustration, and curiosity (see Appendix C.2 for model architecture and dimensional mappings).

 

This feeds into a Relational Policy Graph (RPG) that tracks relationship state—trust, tension, attachment—and models how resistance evolves based on interaction history. The RPG uses memory patterns to detect repeated requests, escalation attempts and avoidance behaviors (see Appendix C.3 for pattern detection algorithms), then selects actions (validate, pushback, escalate, soften, repair) based on weighted goal optimization (see Appendix C.1 for weight formulas and CPT integration).