Every piece published here from August 2026 onwards carries this declaration, at the foot of the page:
Drafted with AI assistance, then checked and revised by the author. The judgements and the errors are the author’s — declaration.
It links here. This page is what it means, and it is the standing policy for everything published on this site from August 2026 onwards.
The material here begins as AI-generated source material, and is then refined adversarially and iteratively across a range of tools — systems built by different companies in different jurisdictions, set against each other and against the author — before being checked against the governance principles argued for on this site. What is finally said is his judgement and his responsibility, on every page, however it was drafted, and the errors are his too.
“Drafted with AI assistance” covers three things: drafting the prose, gathering research, and translating pages into the other languages this site publishes in. It does not cover the argument, the structure, or the judgements. Those are the author’s, and so is the decision to publish.
Source-checking on this site is done by AI agents that did not draft the piece. That is stated plainly because the alternative phrasings — “reviewed”, “verified”, “checked by a reader” — are the wordings one reaches for when one wants the credit of independent human review. A reader would assume a person, and that assumption would be wrong.
What agent checking can establish: whether a cited study exists, whether the figure quoted matches the source, whether the sample size and limits are stated correctly, and whether a claim in the draft is wider than the evidence behind it.
What it cannot establish: whether the argument is sound, whether the sources are the right ones, or whether something important is missing. A checker that did not write the piece is a different reader, not an independent authority.
One of the systems used here began watermarking its text output in August 2026. Neither the author nor the assisting model can verify that mark on a given page.
The reasons are set out at length in The Marks It Leaves, which is an essay about exactly this problem: every mark a machine leaves in text is either removable, deniable, or both, and none of them amounts to proof. That essay’s conclusion is the reason this page exists — the standard is someone saying so, and being held to it.
This section is the one that dates. It lives here, once, so that it can be corrected in one place rather than going stale inside every essay that mentions it.
Where a piece rests on evidence, it carries a sources page listing every study with its source, its sample and its limits — including which come from preprints rather than peer-reviewed work.
Corrections are recorded rather than quietly fixed. Where checking changed the argument, the sources page says so and says where.
Longer pieces may also carry a glossary, a questions-and-answers page, and a slide deck. Not every piece warrants all of them; the sources page is the one that matters when an argument rests on evidence.
This page is the standing policy. It deliberately says nothing about any particular essay.
What checking changed in a given piece is recorded on that piece’s own sources page, under its own heading, because that is where a reader who is checking the argument is already standing. There is no central index of corrections to fall out of date, and no bespoke declaration page per essay.
20 August 2026 — first published. Replaces a longer disclosure block carried inline on individual essays, and an alert-styled banner on blog posts that cited framework instruction identifiers a reader had no way to look up. The facts declared are unchanged; the difference is that they now live at one citable address instead of being restated, and going stale, in every piece.