Michael Darius Eastwood · Independent AI alignment researcher
Published
The ARC/Eden programme is now a suite of twenty-two documents. This note is a reading guide, not a summary. It suggests an order that works for someone starting cold, and it flags where a result is strong, where it is single-lab, and where a headline number is currently under recompute inside the programme itself.
What order should someone read these papers in if they want to understand what is claimed, what is evidenced, and what is still open? The suite mixes formal mathematics, empirical experiments, methodology work and explicitly philosophical essays. Reading it in the wrong order makes it easier to overclaim on someone's behalf than to understand it.
Start with the Executive Summary and the Foundational paper. Those two set the vocabulary. Then read Paper I on the ARC principle and Paper II on the experimental validation, noting the retraction of the earlier unblinded single-model exponent of about 2.24, corrected to about 0.49 across architectures under blinding. Paper III on the alignment scaling problem comes next; its metabolic-scaling headline figures are under recompute and unreconciled inside the programme, and should be treated accordingly. The Paper IV suite is best read as a block: IV-a on baked-in versus computed alignment, IV-b on alignment saturation at low depth, IV-c on the ARC-Align benchmark, and IV-d on the blinding sign-flip. Paper IV-d is the methodological warning that colours everything else. The substrate-and-architecture arc is Paper V (the stewardship gene), Paper VI (the honey architecture), Paper VII (Cauchy unification) and On the Origin of Scaling Laws. Then read the load-bearing and self-correction papers: Paper VIII (two nulls, one confirmation), Paper IX (the honest error catalogue and roadmap), and Paper X (which supersedes the fixed-exponent framing of U = I x R^alpha as the operative safety criterion and installs the beta greater than k co-scaling stability condition; the equation itself and the ARC Bound remain live hypotheses). Paper XI is the convergence register with 19 independently sourced convergences. Alongside these sit HRIH as explicitly speculative philosophy, Paper C on P versus NP as formal mathematics, and the paired Eden Vision and Eden Engineering documents. The Master Table of Contents is the index into the suite rather than a paper you read for content.
Several places in the suite have to be read with care. Paper II carries a retracted unblinded single-model measurement of alpha approximately 2.24; the current cross-architecture estimate is about 0.49 under blinding. The equation and the ARC Bound were not retracted; only the measurement was. Paper III's metabolic-scaling percentages are under recompute and unreconciled and should not be quoted from until they are. Paper VIII's load-bearing test returned null on two of three preregistered experiments and only the gated simulation confirmed the mechanism. Paper IV-d shows the blinding sign-flip: previously published alignment results reversed direction once evaluators were properly blinded. Several results, including IV-d itself and Paper II's blind cross-architecture run, are single-lab and await external replication. Paper C's P versus NP result should be treated as a formal claim requiring community verification. HRIH is philosophy and cannot fail empirically; the appropriate question is whether it is a productive frame. Priority on the quantified sequential-versus-parallel finding is credited to Sharma and Chopra (arXiv:2511.02309, 4 November 2025), which precedes the January 2026 formalisation by roughly two and a half months; the directional thesis appears in the 8 December 2024 manuscript.
Everything in the suite depends on the blind-evaluation methodology of Paper IV-d, and on the Cauchy composition scaffolding of Paper VII. Paper IX and Paper X are the suite's self-correction mechanism: they catalogue what has been empirically retracted (the unblinded 2.24 fit, corrected to about 0.49 under blinding) or narrowed, and update the operative safety framing accordingly. Paper XI is the outward-facing convergence view and should be read against the falsification dashboard rather than in isolation.
Every paper HTML sits under the programme's OSF deposit at DOI 10.17605/OSF.IO/6C5XB. Replication code for the blind evaluation, the sequential scaling measurement and the Love Loop intervention is in the arc-principle-validation repository. The Master Table of Contents links every item; this reading order is a suggested traversal of that same list.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.