Black Box Notes

On opacity, auditability, and the limits of trust in modern AI systems.

Methodology

How the publication audits agentic systems

The publication writes about how others should audit opaque systems. It would be strange not to be obvious about how we audit them ourselves. These pages publish, in detail, the working methods Black Box Notes uses behind every piece in the archive. They double as a reference for working auditors who want to compare notes.

Why publish methodology at all?

The publication treats opacity as a property of the relationship between a system and its observer. A publication that writes about opaque systems while keeping its own methods opaque is, by its own definition, part of the problem. So we publish them.

The pages above are working methods, not aspirational ones. They describe what the editorial team actually does — including the parts of the method we are unhappy with, the questions we cannot yet answer, and the cases where our own audit found nothing useful to publish.

Downloadable audit checklists

Every methodology page ends with a downloadable checklist a working auditor can use as a starting point. The checklists are CC-BY licensed — copy them, modify them, ship them inside your own audit reports. We ask for attribution where you can give it. The downloads live in /checklists/.

How these pages get updated

We revise a methodology page when the publication's working method changes. Revisions are dated at the bottom of each page and listed in a changelog there. Material revisions trigger an editor's note pointing to the change.

Tell us what we got wrong

Working auditors who read these pages and find a method incomplete, mis-described, or plain wrong, please write to editors at blackboxnotes dot com. We have built these pages by reading regulatory output, working audit reports, and our own practitioner contributors' notebooks. We have certainly missed something.