Verification of execution of natural language data analytics requests

Abstract

A system performs verification of results of natural language data analytics requests using a machine learning based language model. The system generates an execution plan comprising ordered data processing steps to produce an analytical result. The system traverses the execution plan in reverse, and for each step, identifies applicable invariants from a seed library or by generating new ones, produces structured query input describing the step and invariants, and uses the language model to generate executable verification code. The verification code is executed on step inputs/outputs to determine invariant satisfaction without exposing raw data to the model. Discrepancies are localized to specific steps, correlated with possible root causes, and presented in a detailed verification report with lineage and metrics. The system provides metadata-driven, privacy-preserving, explainable verification of AI-generated analytics, enabling detection of errors and ensuring correctness across complex multi-source analytical workflows.