Michael Darius Eastwood · Independent AI alignment researcher
Published
Some readers arrive at this material and their first instinct is: this is a machine-generated pile of talking points. The instinct is understandable in a year when a great deal of writing is machine-drafted. It is also structurally irrelevant to whether the claims survive the checks. The evidence chain does not run through the prose.
A model can produce fluent prose that argues anything. It cannot produce a Gmail Message-ID that Google's servers logged in December 2024. It cannot produce an .eml whose SHA-256 was published to a public evidence page eighteen months ago and match today. It cannot retract a number that has already been withdrawn on a dashboard. It cannot fire a preregistered kill-condition. Whatever tool assisted with any specific sentence is irrelevant to the check the reader is asked to run, which is to hash a file and compare a string.
The 8 December 2024 bundle predates any usage pattern of frontier models against the specific structural theses it contains, in the record this programme's author holds. The author does not claim never having used models at all in that period; the specific manuscript's content, its interfaith-governance argument, its embedded-alignment thesis, its recursion equation, its 189,355 words across four attachments, arrived as an authored artefact and was self-emailed at 02:45 UTC. What a model wrote or did not write inside that file is checkable against every subsequent public appearance of the same theses: they are consistent, they are extended in the April 2025 manuscript, they were traceable to the January 2026 book. The record has a shape a one-off model output would not.
Even a maximalist assumption about model authorship of prose does not explain the methodological infrastructure. Kill-conditions on each claim, published in advance. A retraction of alpha 2.24 with a public dashboard entry. A four-layer blinding protocol that produced a sign-flip. A dedup audit that lowered the count from thirty-one to nineteen and reported doing so. A truth gate that lints the corpus and blocks stale numbers. These are not text; they are habits, applied over months, against the author's own optimism. Models can write the prose that describes them. They cannot run them.
Skip the prose. Open the evidence page. Hash the .eml. Resolve any three convergence sources. Read the retraction. The check is against the record, not against the author, and by design it does not care who drafted the sentences on this page. If the check succeeds and the reader still feels the prose is machine-shaped, that judgement is welcome to survive; it is orthogonal to whether the underlying claims are sound. If the check fails at any step, we would like to know. The corpus's response to any failed check is the same as the response that produced its most credible entry, which was the alpha 2.24 retraction: state, correct, keep visible.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.