Controls and Safeguards

Ambiguity and Intent Clarity

• Active Clarification Loops: Solid identifies underspecified questions (e.g., just "sales") and prompts for missing parameters before execution.
• Intent Classification: Solid provides multiple MCP tools so agents can route different need types — glossary lookup, semantic model description, asset metadata, SQL generation — each with a focused purpose.
• Explanation Transparency: Solid provides a reasoning summary of the SQL logic, so the user or agent can verify Solid interpreted their request correctly before acting on the result.

Authorization and Data Security

• Zero-Data Exposure: Solid outputs SQL code and logical explanations. Data retrieval is strictly handled client-side with the user's authenticated credentials — raw data never flows through Solid.
• PII and Sensitive Data Masking: Automated scanning identifies sensitive columns (PII/PHI) to avoid sampling during data collection, using the Azure PII Detection engine. If PII is detected in a sample, that column's data is discarded and not stored.
• Read-Only Enforcement: The engine can only execute SELECT statements — enforced by non-LLM static code analysis, deterministic rather than probabilistic.
• Identity-Aware Execution: SQL is executed on the data warehouse using the end-user's specific permissions. Solid does not replicate or copy access control from source systems; it delegates enforcement to the executing agent, running with end-user credentials.

Hallucination Risk Mitigation

• Governed Schema Scope: The engine cannot reference tables or columns outside the semantic model — no arbitrary table additions or column inventions.
• Automated Self-Correction: A three-attempt correction loop catches execution errors before returning SQL to the agent.
• Latency-Bound Execution: The engine automatically aborts long-running test queries to prevent infrastructure bottlenecks; the timeout is customer-configurable.
• Proactive Model Tuning: Benchmark analysis surfaces actionable recommendations to improve model quality before hallucination-generating edge cases are encountered in production.

See Security Architecture for the platform-level security model these controls sit within, and When Does Solid Execute SQL? for exactly which executor and identity apply at each pipeline stage.