A research framework called GROUND aims to stop language models from inventing metrics, using invalid joins, or bypassing access rules when answering business questions with database queries. It places a governed semantic layer between a user’s request and generated SQL.
The layer supplies approved definitions for metrics and dimensions, permitted join paths, filters, data grain, and row-level security. Before a query runs, validators check those constraints as well as schema and cost rules. The system retries or abstains when a query violates them.
On a synthetic 100-question enterprise-reporting benchmark, GROUND was the only tested approach with no measured hallucinations across six evaluated categories. Schema-only, retrieval-assisted, and semantic-only alternatives produced errors; exact metric definitions without access policy still leaked data. Tests on US vehicle-safety data and an adversarial set across four models also produced zero violations for mechanically enforced filters and row-level security.
The benchmark is limited, and judgment-based behavior such as recognizing an undefined metric remained fallible. The result distinguishes hard controls from prompting: approved definitions improve meaning, but executable policy checks are needed to enforce security. Production teams would still need independent testing against their own schemas, permissions, and adversarial requests.