Interactive Role-Playing Game System Using Dice-Based Mechanics and AI Narration

Tabletop role-playing games such as Dungeons & Dragons depend on a human Game Master to adjudicate rules, resolve player decisions through probabilistic dice mechanics, and weave a coherent narrative across sessions. Automating this role using Large Language Models (LLMs) alone leads to severe reliability problems: unconstrained generative models hallucinate game state, permit physically impossible player actions, and fail to maintain consistent numerical variables over extended play. This paper presents a hybrid software architecture that cleanly separates deterministic game logic from constrained narrative generation. The backend employs a programmatic Python rules engine grounded in a comprehensive static rulebook covering 20 monsters, 18 weapons, 26 spells, 13 NPC archetypes, and 8 puzzle templates to resolve all combat, skill checks, and state mutations with absolute mathematical precision. Six specialized LLM micro-agents—a Validator, Parser, Story Engine, Director, Enemy AI, and Narrator—handle only semantic tasks, each constrained by typed Pydantic schemas and post-generation rulebook validation. A React-based frontend communicates with this backend via both REST endpoints and WebSocket channels, enabling both single-player and real-time multiplayer sessions with up to four concurrent players. Empirical benchmarks over 100 adversarial and valid test prompts demonstrate that the hybrid Validator intercepts 95.0% of rule-violating inputs while incorrectly rejecting only 2.5% of legitimate gameplay actions, compared to an unconstrained LLM baseline that permitted 88% of adversarial inputs. Deterministic rule execution achieves sub- millisecond latency (≈0.19 ms), representing a speedup exceeding 6,000× over equivalent LLM-mediated resolution.