Large Language Models (LLMs) enable open‑ended dialogue in interactive games, yet their nondeterministic outputs often break authorial control, factual consistency, and the intended information‑release order. Detective games are especially vulnerable: premature clues or fabricated details can collapse the player’s logical reasoning. To address this, we introduce a Structured Knowledge Tree architecture combined with a tri‑agent LLM pipeline. The retrieval agent extracts facts permitted by the current narrative node; the generation agent crafts the suspect’s response based on those facts; the verification agent checks the generated text for factual integrity, allowing only verified content to reach the player. This separation guarantees that the virtual suspect discloses information solely when the narrative state authorizes it, enforcing precise epistemic pacing. We evaluated the system in the playable testbed “The Interrogation of Adrian Gale” and conducted a formal user study measuring hallucination reduction, adherence to disclosure sequences, and perceived logical progression. The structured approach cut critical hallucinations by 64.78% and completely prevented premature narrative disclosure. Although the strict mechanical constraints introduced usability trade‑offs, such as forced conversational reveals, the system successfully enforces rigorous epistemic pacing and gives players a clear sense of case advancement.
Review