Publishing Integrity

How we create content

Our blog has two kinds of article, told apart by the byline.

Reports and client work

The same rule covers the rest of what we publish and deliver.

• The published reports open with a box that says which parts a model played or drafted, and what Dr Dan Epstein decided. Dan signs every report before it goes out.
• Client exercises and games run on Anthropic's Claude by default, or on a local model if the client asks. Dan referees and signs every client finding. Where client material goes.
• For any finding where a model reads the evidence, we publish the adjudication statement, a one-page standard: who ruled, in what order they read the evidence, how much rests on judgement, and who checked.

Our editorial process

Every piece we publish, however it starts, goes through the same steps:

1. Editorial review for argument quality, factual accuracy, and audience fit
2. Language and grammar check to make sure it reads clearly and consistently
3. Editing pass to make sure the prose meets our standards
4. Final editorial approval by the human

Our position on AI and content

AI helps with drafting and research. It does not replace editorial judgement.

We badge AI-assisted content so you can tell which is which.

We publish rarely, and heavily. AI helps us cover topics outside the main reports that we would not otherwise have time for. Human review keeps those pieces to the same standard.