I am collaborating with an AI and machine-learning specialist holding a master’s degree to explore how locally hosted large language models can arbitrate games with formal rules and open-ended player actions.
Traditional game engines work well when every possible action can be represented in code. They become more restrictive when players are allowed to describe creative actions in natural language. This project investigates whether a local LLM can interpret those actions, apply the relevant game rules, identify ambiguity, and return a structured ruling without becoming the authoritative source of game state.
The current research uses Ollama to run and compare local models. Areas of evaluation include:
- Accuracy in applying written rules
- Consistency across similar situations
- Structured and machine-readable responses
- Context-window and rule-retrieval strategies
- Latency and hardware requirements
- Hallucination and contradiction detection
- Explainable decision records
- Escalation of uncertain rulings for human review
The architecture separates deterministic and probabilistic responsibilities. The application maintains authoritative state, validates allowed changes, and records outcomes. The LLM acts as an adjudication layer for ambiguous, narrative, or free-form actions. This reduces the risk of allowing a generative model to directly control persistent game state.
The broader goal is to determine whether the same design pattern can support other specialized adjudication systems where decisions must combine written policies, contextual information, structured outputs, and an auditable explanation.