Large language models (LLMs) enable flexible conversational interfaces, but remain difficult to deploy in structured, staged tasks that require controlled progression, bounded information disclosure, and task validity. Prompt-based control is inherently probabilistic and has been shown to degrade under multi-turn interaction, leading to premature solution disclosure, stage skipping, loss of state coherence, and constraint violations. This study presents a constraint-driven conversational architecture that separates probabilistic language generation from deterministic task governance. An external control layer manages dialogue state, stage transitions, and constraint enforcement, while a simulation layer represents task logic independently of the LLM. We instantiate the architecture in the context of customer discovery tasks to illustrate how staged processes and bounded disclosure can be operationalized without embedding task logic directly into prompts. This work focuses on architectural design and control mechanisms rather than outcome evaluation, offering a reusable architectural pattern for LLM-driven conversational systems that must preserve staged progression, enforce constraints, and prevent premature disclosure during multi-turn interaction.
Ilagan et al. (Thu,) studied this question.