Fernando Reitich

LLM protocols can actively *harm* accuracy through "corruption," and this paper provides a way to measure and mitigate this effect, turning opaque pipelines into auditable modules.