This paper proposes a truncation-first controlled Hidden Markov Model (HMM) framework for online estimation of bounded, uncertainty-aware proxy scores from user inputs in intervention-aware chat systems. The target is capability improvement for adaptive AI behavior (e.g., response adaptation and intervention timing) rather than psychological or medical diagnosis; the outputs are treated as head-conditional proxy-score distributions rather than intrinsic measurements of human traits. Core components include: (i) a controlled latent regime chain whose transitions depend on executed interventions and an exogenous semantic-context tracker summary, (ii) ordinal proxy-score heads (bounded discrete scores) with uncertainty propagation via tracker marginalization, and (iii) an explicit two-stage response process that separates prompting from responding and clarifies missingness assumptions. The framework provides leakage-safe pre-turn prediction and post-turn nowcasting, with a formal definition of prediction leakage and a sufficient condition for leakage-free semantics. Learning is formulated via (batch) EM and fixed-lag online EM, making approximation loci explicit (tracker representation and numerical integration). The model also supports optional extensions for partial personalization via hierarchical user random effects and for open-ended actions via bounded continuous action descriptors, while preserving safety-compatible transition regularization. A distinctive feature is the model-level safety gate integrated into the action pipeline: uncertainty and calibrated risk diagnostics are mapped to executed actions under explicit governance constraints. The paper delineates identifiability boundaries, distinguishes latent-MAR response modeling from an optional self-masked MNAR sensitivity extension, and specifies enforceable privacy/legal/IP controls and prohibited high-stakes uses.
Dai et al. (Fri,) studied this question.