PL-PRISM: A White-box Auditing Framework for Dynamic Persona and Emotional Alignment in LLM Agents

Long-lived companion LLM agents exhibit three structural failure modes in long-horizon interaction: (i) loss of longitudinal state—stateless per-turn inference discards the user’s prior settings and constraints, so each session re-initializes and affective continuity is lost; (ii) unobservable state and hallucinated assumptions—because the agent is a black box, it frequently presumes unverified state (e.g., assuming a user action is already done) while offering the user no window into its internal context; and (iii) overt affective intervention and loss of autonomy— some agents deploy over-emotional, coercive phrasing to steer the user’s behaviour, with little respect for interaction boundaries or safety guardrails. We propose PL-PRISM, presented here as a V1 conceptual and architectural framework. PL PRISM reframes the agent as a white-box dynamical system. It does not purport to solve general long-term memory; instead it makes the affective state persistent and observable: a Pleasure–Arousal–Dominance (PAD) substrate evolves across turns under half-life decay, a zero-dependency JSON telemetry stream exposes the agent’s internal affective trajectory under physical isolation, and out-of-character (OOC) anchor guards together with agent-side do-not-disturb and tool-safety gating keep expressive output within a safe boundary. We present the V1.0 design and invariants; no human-subject evaluation is reported. ---**Code & Full Framework Release:**The lightweight open-source version is available at Gitee: [https://gitee.com/zsy060116/pl-prism](https://gitee.com/zsy060116/pl-prism)The complete theoretical framework (encrypted release) can be downloaded from Gitee Releases: [https://gitee.com/zsy060116/pl-prism/releases](https://gitee.com/zsy060116/pl-prism/releases)