From Prompt to State: Verifiable Grounding and Operative Replay for LLM-Mediated Control

Summary

When a language model controls software, its words must be interpreted as commands before they can change the application. This paper records the connection between the model’s response, the derived command, the applicable rules and the resulting state. Implemented in BatLLM, the approach makes it possible to check and replay a recorded decision’s computational consequences without needing the model to generate the same response again.