Why LLM Decisions Should Be Deterministic
Why LLM-driven systems need deterministic decision layers, and how auditability, not just consistency, matters for engineering teams.
An engineer argues that decisions in LLM-based systems should be made by deterministic code, not the model itself. Initially framed as a consistency fix, the deeper value turns out to be auditability: a third party should be able to reproduce a decision from the implementation rather than trusting the model's after-the-fact explanation. In a real-world alert classifier, model output is validated against a fixed enum via a simple Python function; anything that doesn't match falls back to 'unknown.' This makes every classification reproducible purely from source code. Regulatory frameworks like the EU AI Act and NIST's AI RMF enforce the same principle from the opposite direction, requiring logs detailed enough to reconstruct high-risk system behavior. The author notes that when a model explains its own output, that explanation is a new probabilistic inference, not a genuine causal account. The conclusion: narration remains useful, but it should never be the source of a decision, and engineers must clearly separate which outputs truly need to be reproducible.