Policy as Code, Graded Twice: Keeping an AI-Use Policy and a Control Self-Assessment True to a Running System
A methods preprint by Jake Lawrence, published under CC BY 4.0 and not peer reviewed. It describes two small mechanisms that tie a written policy to a running system with ordinary software tests: a drift guard, which fails when a file or text that a policy rule cites as its enforcement is gone, and a two-lens grade, which grades each control once for design and once for operation. It reports what both did on one platform operated by one person, including a replay of 59 days of history that shows five stale citations repaired and the cases the guard missed. The two artifacts it measures are self-assessments by their author. Neither is an audit or a compliance claim, and nothing here is legal advice. Version 1.0. A preprint: not peer reviewed. The PDF is rendered from the Markdown file beside it, which is the source of truth. Datasheet, attribution and known defects: DATASHEET.md in this record. Checksums: manifest.json. Live page: https://www.jakelawrence.xyz/research.
Authors
- Jake Lawrence
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24430329
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint