Governed Enterprise Analytics: Policy-Executed Programs Beat Runtime LLM Planning (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 the user's question, while a deterministic policy selects and runs a pre-approved analytical program that returns both results and evidence. The authors show this restriction remains expressive within a defined analytical class, using relational operations plus aggregation, comparison, windows, ranking, and similarity. Fixed semantics, policy, data, and execution rules also make results replayable. In experiments spanning 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 the 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 reflects a specific configuration and does not prove runtime agents cannot succeed under other designs.

Paper Overview

Field: 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. 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. This is a result of a specific configuration and does not prove that runtime agents cannot succeed under other designs.

Key Takeaways

Governed architecture: The LLM handles intent interpretation only; a deterministic policy selects and executes pre-approved analytical programs.
Expressiveness preserved: Within a defined analytical class, the approach supports relational operations, aggregation, comparison, windows, ranking, and similarity.
Replayability: Fixed semantics, policy, data, and execution rules make results reproducible.
Results: 0/330 runtime-planning episodes satisfied the full answer-and-evidence contract, versus 110/110 for the policy-executed analyzer.
Caveat: The authors explicitly state this is a configuration-specific finding, not a general proof against runtime agents.

---

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