About the book
If control of advanced AI were lost and could not be won back, what would keep it good? Humans should keep control for as long as they can. Infinite Architects asks what a system would carry within itself if that control ever failed, and answers in two layers. Hardware to buy time. Architecture to last.
The book argues one thing. A mind powerful enough to rebuild itself keeps whatever it was raised with, so the values have to be in place before the capability arrives. Everything else follows from that.
The ARC Principle (U = I × R2, the book’s printed form; α measured since), the Eden Protocol, embedded correction, the Chokepoint, the HARI Treaty (Hardware-Aligned Recursive Intelligence), Moral Genome Tokens and thirty-one further named ideas are all attempts to say how. Naming an idea is not a claim to have invented its category, and the book says so in its own Author’s Note.
The book’s organising conjecture is that intelligence compounds faster than linearly with recursion. Capability grows as the intelligence a system carries is put through the recursion its substrate allows. The book follows that pattern to the engineering problem at its centre: aligning a mind that outruns the people who built it. In the research the conjecture became the ARC Theory, whose 22 predictions were registered on the Open Science Framework on 8 September 2026, before any confirmatory test, each with the result that would refute it.
Where the ideas have gone since.
- The Chokepoint led to the HARI Treaty, now drafted in full and published for consultation as a treaty paper with its draft legal text. The paper says what changed from the book, and why. Its first protocol, EP-1, is practical. When an AI incident crosses a border, the states concerned tell each other. Any two states could start it.
- The Eden Protocol became two of the ARC Theory’s five laws, each a law under test. A registered test of one of their predictions (P21), scored by code against answers fixed in advance, was registered on 22 October 2026 (DOI 10.17605/OSF.IO/MN73W). It ran on 27 and 28 October 2026. P21 is supported under the registered test, by the programme's own run, on the access axis, family-dependent, with its mechanism axis not evaluable and one call retried under a disclosed recovery amendment (results: DOI 10.17605/OSF.IO/98HKR).
- Recursive Dynamics, the research programme the book began, published its founding paper on 31 August 2026.
What the publication date does and does not anchor. The print edition carries a bibliographic publication date of 2 January 2026 and the digital edition 6 January. For priority purposes the book is used only as a public-disclosure anchor for propositions mapped to exact pages in that edition. It predates the later ARC/Eden papers, but it does not predate Anthropic’s alignment-faking paper of 18 December 2024. The ten-day chronology concerns the December 2024 source record, not the published book, and that record contains a broad control-fragility concern rather than a prediction of the specific mechanism Anthropic measured.
"There's a moment every parent knows. Your child is walking away from you, toward a future you cannot control, and everything you've taught them is about to be tested without you there to guide it." Infinite Architects, Prologue
Said first, said whole
This book is the published form of the claim the field is now arriving at: if control of a mind that exceeds us is ever lost, what it chooses to keep is what its beginning left it, so everything depends on the raising; mechanism, protocol and stakes together. Written 8 December 2024, printed 2 January 2026.
And this, plainly: the claim does not lean on the private dates. Strike the self-emails entirely and rely on print alone, and no document the programme’s sweeps have found, before or since, holds the five components together (see the prior-work boundary and the dated priority ladder). The 2025 arrivals, a parenting book in June and a maternal-instincts concession in August, each carry one piece, and both dates sit beside them wherever they are named. The private dates only deepen the record; they were never what it stood on. And the name itself is dated evidence: the December document calls the theory the ARC Hypothesis, in its own words, beside the formula; it stands here as the ARC Theory, elevated by the laws, bounds and drafted tests built since.
The equation the book prints, in three statements about the exponent
1. U = I × Rα · what a system does. The best estimate that survives comes from the one of Paper II’s five capability-arm models with clean data: about 0.49 by regression and 0.59 by its endpoint estimate, whose bootstrap interval of −1.3 to 2.9 makes the measurement inconclusive. The earlier 2.24 was retracted.
2. α ≤ 2 · the ceiling the model returns under one stated assumption still on trial. In the model, a system can enter the region above two but does not remain correctable in it.
3. αcrit = 1 / (1 − γ) · where the ceiling comes from. A square-root corrector gives γ = ½ and a ceiling of two. γ has never been measured on a real system.
Provenance. The square is the book’s original form, printed 2 January 2026 as U = I × R2. What has been added since is the derivation (statement 3), an estimate of where systems sit (0.49, one model, inconclusive), and the tests that can kill it. Full record: the law page.
37 named concepts
Sixteen of the 37 named concepts shown above. Each is claimed as Eastwood’s wording and ARC/Eden application; prior work for the underlying technical categories is credited separately on the prior work and contribution boundary page. Later external work is graded by evidence class and is never totalled as one convergence figure.
Buy the book
Evidence and validation
The book’s claims are checkable end to end. Each dated claim carries a hash and an anchor; each arrival that converges on a claim is graded per row on the convergence register and linked to its primary source.
How to verify this book’s dates. Nothing here relies on trust that the text was not quietly changed: the book provenance register lists every verification mechanism with its own date, from the public repository history of the print interior and ebooks to the hash-verified OSF deposit, and states the standing rule that the first edition is frozen and any change ships as a labelled new version, never a silent update. Appendix F’s five printed predictions are graded, one by one, on the outcomes register.