State-Aware Safety-Gated Controlled HMM for Online User-Input Signal Estimation in Intervention-Aware Dialogue Agents

Abstract

This preprint develops a safety-gated controlled hidden Markov model for online estimation of bounded proxy user-state signals in intervention-aware dialogue agents under uncertainty and governance constraints, yielding leakage-safe prediction and risk-aware adaptive actions. It separates prompting and response mechanisms, combines pre-turn prediction with post-turn nowcasting, and supports EM-based learning with explicit missingness and identifiability assumptions for auditable AI operation.

Keywords

• controlled hidden Markov model
• intervention-aware dialogue agents
• online state estimation
• uncertainty-aware proxy scores
• safety-gated action policy
• leakage-safe prediction
• fixed-lag online EM
• ordinal bounded scores
• response missingness modeling
• identifiability boundaries
• adaptive AI safety
• auditable governance

Identifiers and source records

• DOI
• Works entry
• ORCID