Preprint proposes governed layer for enterprise AI analytics

A preprint introduces GROUND, a framework intended to constrain LLM-generated enterprise analytics to approved business definitions and validate SQL before execution. It checks metrics, joins, grain, filters, row-level security and cost rules, retrying or abstaining after violations. On a 100-question synthetic reporting benchmark, the authors report no measured hallucinations in six evaluated categories; they also tested the approach on NHTSA vehicle-safety data and an adversarial set across four models. The work remains a preprint rather than a production deployment.