SmartVerify glossary
Prompt-to-execution drift
Definition
Prompt-to-execution drift is the difference between what an AI agent was asked to do and the database statement that actually ran.
Scope
Measuring the drift requires two records: the tool call that represents the agent's request and the database execution record. A database execution record remains useful on its own, but it cannot establish the difference from the original request without the corresponding tool-call record.
Examples
- An agent is asked for one customer record, but the resulting statement returns every customer in a region.
- A tool call requests a bounded export, but the executed statement omits the expected limit.
- A request names a non-sensitive field, but the executed statement reads additional sensitive columns.
Current availability
In design partner buildSmartVerify compares the two records when the deployment topology exposes the agent-to-tool connection. The inline edge worker in front of PostgreSQL is always present. MCP and JSON-RPC observation and PostgreSQL execution records are available now where those connections pass through the customer network path. Vendor-hosted tool servers may leave no customer-side connection for the tool-call listener. The Platform coverage grid is the source of truth for availability.
Related technical detail: Reconciliation.