Intel / article

Code-as-Auditor: Executable Compliance Reasoning via Regulation-to-Code

Source: arxiv.org
Author: Jisoo Kim, Taeyoon Kwack, Jinwoo Jang, Woo Kyung Kim, Honguk Woo
Date: Published Sep 18, 2026

Context

Code-as-Auditor aims to make an AI compliance assessment traceable from evidence to a regulatory rule. It converts regulatory information into checklists and executable decision trees, expressing conditions as code that can be inspected. For each checklist item, it then asks questions about the case and about alternative circumstances that could affect the conclusion. Framework description.

The proposed benefit is a visible chain connecting case evidence, rule application and the final assessment. The authors report more accurate, better-supported evaluations in privacy and data-protection scenarios, but the abstract gives no numerical improvement. The supported contribution is the structured assessment process and its reported , rather than a demonstrated guarantee of legal correctness. Reported scope.

Key takeaways · AI-distilled

Converts regulations into executable decision trees and checklists, encoding legal conditions as interpretable code rather than free-text prompts.
At , each checklist item expands into factual and counterfactual questions so the model reasons against case-specific evidence, not just static rules.
A self-verification loop checks the generated code's logical consistency, improving traceability of how each compliance decision was reached.
In tests on privacy and data-protection scenarios, the framework delivered more accurate, evidence-backed compliance evaluations than ungrounded outputs, the authors report.

Terms in this piece · Glossary

eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source arxiv.org