Governed Enterprise Analytics: LLMs Interpret, Deterministic Policy Executes (arXiv 2509.00002)

Forum topic · 小凯 · 2026-09-06

Summary

This arXiv paper (2509.00002) from MasterControl AI Lab studies a governed approach to enterprise analytics in which a language model interprets user intent while a deterministic policy selects and executes a pre-approved analytical program that returns both results and evidence. The authors show that this restriction remains expressive within a defined analytical class, covering relational operations plus aggregation, comparison, windows, ranking, and similarity, and that fixed semantics, policy, data, and execution rules make results replayable. In experiments across 440 runs, three 8B models generated SQL and selected tools at runtime, while Qwen3-8B only interpreted intent and policy executed approved programs. None of 330 runtime-planning episodes matched the full answer-and-evidence contract across all test datasets, whereas the policy-executed analyzer matched 110 of 110. The authors note this is a result of a specific configuration and does not prove runtime agents cannot succeed under other designs.

Overview

Research area: AI/ML
Authors: MasterControl AI Lab
Published: 2026-09-06
arXiv: 2509.00002

Abstract

We study a governed approach to enterprise analytics: a language model interprets the question, while deterministic policy selects and runs a pre-approved analytical program that returns both results and evidence. We show that this restriction can remain expressive within a defined analytical class, using relational operations plus aggregation, comparison, windows, ranking, and similarity. Fixed meaning, policy, data, and execution rules also make results replayable.

Key Findings

Across 440 runs, three 8B models generated SQL and selected tools at runtime, while Qwen3-8B interpreted intent only and policy executed the approved program.
None of 330 runtime-planning episodes matched the full answer-and-evidence contract across all test datasets.
The policy-executed analyzer matched 110 of 110 episodes.
The authors emphasize this is a result of a specific configuration and does not prove that runtime agents cannot succeed under other designs.

---

*Auto-collected on 2026-09-06.*