The Model Is a Witness

Short summary

The author argues that language models should recognize events but never define them in money-touching pipelines. After an LLM nearly booked non-existent revenue by confusing a DocuSign request with a signed contract, a deterministic prefix check was inserted as a guardrail. Applied to a medical-billing denial agent, anchoring LLM recommendations to a reference database raised action accuracy from 36.7% to 90.7%. The pattern: deterministic layers make the rules, the model gives testimony.

• •LLMs can perceive events but should not define business-critical outcomes
• •Deterministic anchoring raised denial-agent action accuracy from 36.7% to 90.7%
• •Pattern: deterministic layer makes the law, the model serves as witness

Generated with AI, which can make mistakes.