About

Purpose

Work on verifying claims about AI hardware and software is spread across papers, preprints, technical notes and project pages, with no shared, maintained map. This site aims to be that map: what each mechanism can establish, how mature it is for verification use, what blocks it, and how mechanisms depend on one another. This version holds 25 mechanisms, 17 implementations and 287 sources.

Audience

• Technical newcomers getting up to speed on the field.
• Policy and funding readers who need the lay of the land.
• AI research agents, which can read the machine-readable export.

What it is not

• Not advocacy for any agreement or policy.
• Not a final verdict on maturity. Readiness levels are editorial assessments under a published rubric, and are open to correction.

Maintenance

The site is a project of the Strategic Stability Alliance. Dan Girshovich maintains it on a best-effort basis, with help from AI agents. Agents drafted the records from public sources and checked them against their citations. These checks are not independent review by domain experts. Corrections from people who know the work are welcome through Suggest an edit.

Credits

The site's claim examples overlap with the broader menu in James Petrie's publicly shared Research Priorities for AI Compute Verification, especially its “What Would Need to Be Verified” section. The site's selection and grouping are an editorial synthesis.

The bibliography was seeded from Will Hodgkins' AI Workload Verification Papers, which is licensed CC BY 4.0 and builds on earlier reading lists by Mauricio Baker and James Petrie. The site's structure borrows from the rubric-based maturity stages of L2BEAT, the versioned, git-based data of MITRE ATLAS, and the neutrality norms of the Stanford Encyclopedia of Philosophy and Wikipedia.

Content is licensed CC BY 4.0, and code is licensed MIT.