Why One Developer Banned LLMs From Making Final Verdicts in an AI Audit Tool

A developer building an AI audit system for fitness training plans deliberately excluded large language models from the final judgment step, citing concerns over non-determinism and lack of auditability. The system checks LLM-generated workout plans against published sports-medicine guidelines, returning a pass, flagged, or rejected verdict tied to a specific rule and source. Instead of letting an LLM decide compliance, the architecture uses the model only to extract structured data from free-form text, then hands that data to deterministic Python code for rule evaluation. This two-zone design aims to ensure verdicts are reproducible and verifiable, since an LLM-produced verdict cannot be reliably diffed or traced to a specific threshold. A real bug in the project illustrated the risk: the model correctly extracted a user's age but failed to flag them as a minor because that logical inference had been left on the uncertain side of the boundary rather than enforced in code.