GROUND records zero violations across six hallucination categories in 100-question enterprise benchmark
GROUND constrains LLM analytics to a governed semantic layer and records zero hallucinations across six categories where ungoverned baselines fail. The framework demonstrates that metric fidelity alone cannot substitute for enforceable access policies. Results replicated on synthetic and real NHTSA data across multiple models.
GROUND supplies approved business definitions to the model, binds intent to those definitions, and enforces schema, metric, join, grain, filter, security, and cost constraints on every generated query. On violation the system retries or abstains. Direct schema-only text-to-SQL, schema-RAG, and semantic-only baselines each produced violations in multiple categories, including row-level security leaks on many questions.
The 100-question synthetic set and an independent NHTSA vehicle-safety replication both showed semantic-only grounding with exact metric definitions still permitted data leakage when access policies were absent. Enforced guarantees held with zero violations across four models from three providers; only judgment-dependent behaviors such as refusal of undefined metrics remained fallible. These controls directly reduce erroneous data-driven decisions by blocking unsafe or incorrect analytics at generation time.
Operational deployment therefore requires a single source of governed definitions rather than post-hoc prompt engineering or retrieval alone. Future production systems will need to treat row-level security and cost bounds as first-class validation predicates equivalent to SQL syntax checks.
Snowflake: production semantic-layer governance layers will cover 30 percent of customer warehouses with LLM query interfaces by end of 2027
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.26157)
- [2]Supporting Source(https://arxiv.org/abs/2305.03130)