Michael Darius Eastwood · Independent AI alignment researcher
Published
A scrupulous reader inspects the reference lists and points out that the programme repeatedly cites its own manuscripts, its own papers, its own timestamp records. This is a fair thing to notice. It is also a different thing from what the objection usually implies, which is that the programme is standing in a circle handing itself credibility. Two distinctions are worth drawing.
When the corpus cites the 8 December 2024 manuscript, it is not appealing to the manuscript for authority; it is pointing at an artefact a stranger can download, hash and open. The citation is a pointer to a verifiable file, not an appeal to a prior conclusion. This is the same shape as a physics paper citing its own dataset: the citation is not asking to be trusted, it is directing the reader to the object they are being invited to check. If the manuscript did not exist, the citation would collapse the moment anyone tried to open it. The whole architecture rests on that failure being loud and public.
The nineteen convergence rows each cite the source that carries the claim: arXiv:2412.14093 for the Anthropic result, Nature 2024 for Willow, PRL 136 057201 for the NYU time crystal, arXiv:2606.28639 for the undecidability proof, arXiv:2506.15093 for FlexHEG, and so on. When the programme reports what Anthropic found, it does not cite the programme; it cites Anthropic. This is the specific failure mode the truth gate exists to prevent. A version of the corpus that made external claims by citing internal derivatives would fail its first audit and did fail an early one, before the current pattern locked in.
Three things and only three. Its own methodology, because that is the thing the corpus authored, and describing method by citing method is not circular. Its own timestamps, because those artefacts are the primary sources for their own existence. Its own retractions, because a retracted claim's history is preserved for exactly the reasons discussed elsewhere in this series. Everything else, the reader is directed to a source that is not the programme, and the source can be opened and read without the programme being trusted for a single sentence.
The sharper version says: even if each individual cite is legitimate, the overall effect is a self-reinforcing narrative in which the programme's centrality is quietly asserted by repetition. Fair. The corrective is that the centrality claim, that the framework tracks something real about the territory, is exactly the claim the falsification dashboard and the register are structured to allow a reader to reject. A hostile reviewer who dislikes the whole architecture is invited to work through the nineteen rows and produce the row that fails. The self-cite pattern is not defending anything the external sources are not already carrying. Take the sources away and there is nothing left; and every source is external and checkable. Which is the point.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.