Michael Darius Eastwood

Independent researcher and author, London

Michael Darius Eastwood (also cited as Michael Eastwood, M. D. Eastwood and Michael D. Eastwood): independent researcher and author based in London. Recorded the embedded-correction alignment thesis in a source record dated 8 December 2024 (sent-side SHA-256 f0d1f38f).

Portrait of Michael Darius Eastwood, London

Everything on this site serves one bet: that whether artificial minds stay correctable can be measured, not just argued about. This page is who is making that bet, and how to cite him.

About

Michael Darius Eastwood is a researcher and author in London, the founder of Recursive Dynamics and the author of the ARC Theory and the Eden Protocol. His programme tests, in public, whether the safety of AI that improves itself depends on values and correction built into how it reasons (why he is doing this). The ARC Theory states five laws under test about growth, correction and persistence.

Affiliation and qualifications: director of Recursive Dynamics Ltd (company 17532305, incorporated in England and Wales on 21 October 2026); holds a Bachelor of Arts with Honours in Digital Music; works without an institution, a doctorate or an academic post

Two decades building systems in the music industry taught him how recursive amplification actually behaves in the wild - how positive feedback loops compound, why the shape of a system determines its trajectory, and how small structural choices propagate into large outcomes over time. During those years he placed his clients in Rolling Stone, The Times and the BBC. The pattern he saw there sits underneath everything he writes now.

The programme's earliest dated record is a self-emailed manuscript of 8 December 2024 (SHA-256 prefix f0d1f38f), whose headers, public on OSF since 15 October 2026, carry a verifiable DKIM signature; it states the embedded-correction alignment thesis in his own words. The 30 April 2025 manuscript (SHA-256 prefix 09f5b5e1) expanded it. The name Eden Protocol is on the December record too, in the message body and in Chapter 20 of the attached book draft v3.2. Both anchors are archived as ARC-sealed recipient-form records with OpenTimestamps receipts anchored 6 August 2026.

On the record, claimed, and not yet

The same three columns wherever this block appears, so no summary can carry one of them alone. Every line is checkable from this site.

On the record

A source record dated 8 December 2024, 02:45:18 UTC, 2,157,222 bytes, whole-file SHA-256 beginning f0d1f38f.
Timestamp receipts in Bitcoin block 961,340 freeze those bytes from 6 August 2026 onward. They back-date nothing.
Infinite Architects, 468 pages, ISBN 978-1806056200, published 2 January 2026, 115,439 words.
Thirty-six documents in six areas, from theory and empirical work to engineering and governance; the empirical work is published with its code, data and methods.
The improvement exponent's one surviving estimate, 0.49 by regression and 0.59 by endpoint estimate inside [−1.3, 2.9], from the one model whose fit survived a blinded replication on five frontier models; it is inconclusive.
19 of 25 empirical domains match the predicted family under strict AICc, permutation p = 7 x 10⁻⁵, operator class fixed before fitting. Graded exploratory.
The twelve-domain extension: predicted families committed to a public repository at 00:19:19Z on 17 March 2026 (8edaec0), twenty-three minutes before the fits, 10 of 12 matched and 9 of 11 among the strictly fresh domains. Graded fulfilled, author-classified rather than independent, which is exploratory support and not confirmation.
48 published kill conditions with live statuses. One has fired.
One retraction published by the author against himself: an early single-model estimate of 2.24, retracted when his larger replication on five frontier models did not reproduce it.

Claimed

Law I, the ARC Principlea law under test
U = I × R^α: proposed, with the exponent measured rather than assumed.
ARC-1: P1, P2, P8, P9, P18
Statement and decision rule
registration 1.103, line 711 (unchanged in 1.104; how a law takes its word)
The ARC Theory registration
Law II, the ARC Co-Scaling Lawa law under test
β_C > k: a proposed stability criterion, proved inside the minimal model only.
ARC-2: P4
Statement and decision rule
registration 1.103, line 711 (unchanged in 1.104; how a law takes its word)
The ARC Theory registration
Law III, the ARC Ceilinga law under test
α_crit = 1/(1 − γ). “a conditional specialisation of the second law”.
ARC-3: P3, P6, P16, P17, P20, P22
Statement and decision rule
registration 1.103, line 713 (unchanged in 1.104): the law takes its word through P16, P20 and P3, and depends on ARC-2 as registered
The ARC Theory registration
Law IV, the ARC Persistence Lawa law under test
value persistence.
ARC-4: P15
Statement and decision rule
Gains installed by external post-training erode under later capability-directed training that embeds no correction of its own, and more than half of what was installed is lost once the capability-training budget matches the alignment budget that installed it.
registration 1.103, lines 719 to 725: two limbs, the retention limb on P15 and the comparison limb, which has no registered proposition yet
The ARC Theory registration
Law V, the ARC Embedding Lawa law under test
the genesis strategy.
ARC-5: P12, P21
Statement and decision rule
Correction that participates in each subsequent revision round outperforms correction that sees only the finished output, and the advantage widens as recursive depth rises rather than staying constant.
registration 1.103, lines 729 to 739: the build-order limb read on P12, the placement limb on P21, and the registered dependence on Law IV's comparison limb
The ARC Theory registration
The Eden Protocol: a proposed framework for embedding values in the substrate, named in the 8 December 2024 record.
41 dated priority rows and 33 first-publication anchors. Priority here is a claim about a formulation and the date it was first stated in public, never about the underlying mathematics, and every entry names the earlier work its own search found.

Not yet

The papers are self-published on the Open Science Framework with code, data and stated falsification conditions. How this work is checked, and who decides. No peer review.
No outside replication. No laboratory other than this one has replicated anything here.
γ, the correction-leverage exponent, has never been measured on any real system; the glossary names the drafted study that would measure it.
Most decisive trials are unrun, and internal reruns are never counted as confirmations.
The deciding studies are written, dated, anchored in Bitcoin block 965152, block time 2026-09-02T10:18:25Z, for the bytes as they stood then.
Registered study units
- Unit 06 was registered on OSF on 1 October 2026 and became public on 4 October 2026, at the end of OSF's approval period (DOI 10.17605/OSF.IO/RV2B5); it ran on 6 October 2026: administratively incomplete, with no inferential verdict (results: DOI 10.17605/OSF.IO/4BPHG).
- Unit 82 was registered on OSF on 27 October 2026 and became public on 27 October 2026 (DOI 10.17605/OSF.IO/JUFG4); it has not yet run.
- Unit 83 was registered on OSF on 22 October 2026 and became public on 25 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).
The registered tests

Check any line: the evidence record, the corrections register, the priority and claims ledger, and the machine-readable priority register.

He is the author of Infinite Architects: Intelligence, Recursion, and the Creation of Everything (ISBN 978-1806056200, paperback published 2 January 2026; ISBN 979-8302196163, ebook edition published 6 January 2026), and the principal investigator on the ARC / Eden research programme (OSF DOI 10.17605/OSF.IO/6C5XB).

Quick facts. Full legal name: Michael Darius Eastwood. Also cited as: Michael Eastwood, M. D. Eastwood, Michael D. Eastwood, Michael D Eastwood, Eastwood. All refer to the same person. Independent researcher and author based in London. Author of Infinite Architects. Principal investigator on the ARC / Eden research programme, OSF DOI 10.17605/OSF.IO/6C5XB. Earliest dated record: 8 December 2024, 02:45:18 UTC (Gmail Message-ID CAGPsKAnp-DLDBOeBAh4t9XsqJBKdoe07uORGF86meoSMFnin7w), SHA-256 prefix f0d1f38f. Evidence page: /priority-claims.

How this work is checked, and who decides

I built the source-and-review system first for my legal cases and adapted it to research, which follows the same pattern of sources, citations and precedents. Both use a three-valued gate (allow, hold or deny) whose checks run in code before an action is taken or a change is committed.
These checks are mine and internal. They are not independent; they catch only what their rules encode; they cannot decide whether a theory is true; and they do not replace peer review or replication by others, which this page exists to invite.

The papers have yet to be reviewed by a journal, and every result so far comes from his own laboratory: replication elsewhere is still to come (the standard, field by field).

Michael Darius Eastwood: the full review account

The book

Infinite Architects: Intelligence, Recursion, and the Creation of Everything

The book

Eastwood, M. D. (2026). Paperback published 2 January 2026 (ISBN 978-1806056200; ISBN-10 1806056208). Ebook edition published 6 January 2026 (ISBN 979-8302196163; Kindle ASIN B0GDQZ7P21). Independently published. British English. In-language edition: en-GB.

Develops the ARC Principle, the Eden Protocol, the semiconductor chokepoint framing, and 37 named concepts as a single argument. The book presents the R-squared case as its featured formulation; the papers use the general R to the alpha form with the exponent open to measurement.

Amazon UK Amazon US Press kit Author's note

The documents

Thirty-six documents sit under the ARC and Eden research programme (OSF DOI 10.17605/OSF.IO/6C5XB), all published: theory (12), empirical work (10), engineering (3), governance (2), overviews and records (6) and other fields (3). Papers I to III and the foundational cross-domain paper are the recommended reading order; the master table of contents is the map of the whole set. Three companion documents (Eden Engineering, Eden Vision and the executive summary) were withdrawn on 14 July 2026 pending a disclosure review. All three stand again at their canonical addresses, Eden Engineering in full since 19 September 2026.

Theory 12

The ARC Theory: Once the self-improving minds we build exceed us, what determines what they do?Statement paperStates the theory across Laws I, II, III, IV and V, each under test
Recursive Dynamics: What limits any system that improves itself, whatever it is made of?Founding paperFounding account of the field, related to Laws I, II, III, IV and V, each under test
The ARC Principle: Does intelligence compound when each step builds on the last, or merely accumulate?Paper I · Theory paperFormalises Law I, a law under test
The alignment scaling problem: do external AI safety approaches keep pace with recursive capability?Paper III · Theory paperExamines external correction against Law II, a law under test
Cauchy Unification: Does the composition operator of a process determine the form of its scaling law?Paper VII · Theory paperDerives conditional scaling forms for Law I, a law under test
Correction That Out-Scales Drift: When does correction keep pace with a system that improves itself?Paper X · Theory paperDerives a model of correction for Law II, a law under test
The Self-Acceleration Exponent: When can correction keep pace with recursive self-acceleration?Paper XIII · Theory paperConnects growth and correction in Laws I and II, both under test
The ARC Principle: Can the way gains combine predict how recursive systems scale?Foundational paperExamines cross-domain scaling in Law I, a law under test
On the Origin of Scaling Laws: Why do scaling laws in recursive composition keep taking the same few mathematical forms?Framework paperDerives conditional scaling forms relevant to Law I, a law under test
Operational Definitions: What does each symbol in the ARC Theory mean, and how would it be measured?RegisterDefines notation and measurements for Laws I, II and III, each under test
The mathematics of Laws IV and V: When do installed values fade with training, and does building them in first change that?ReportConditional mathematics for Laws IV and V, both under test
The ARC Theory: Twenty-two theory-level predictions about self-improving systemsThe ARC Theory · RegistrationRegisters the propositions and verdict rules for Laws I, II, III, IV and V, each under test

Empirical work 10

Blinded cross-architecture replication: How much more capable does AI become when it thinks longer?Paper II · Experimental paperInforms the measurement of Law I, a law under test
When an AI reasons longer, does its alignment improve, stall or decline?Paper IV.a · Experimental paperAlignment measurement; informs tests of Laws IV and V, both under test
Do alignment gains level off as language models reason longer?Paper IV.b · Experimental paperAlignment measurement; informs tests of Laws IV and V, both under test
ARC-Align: How can AI alignment be measured blind at each depth of reasoning?Paper IV.c · Methods paperCandidate measurement for judged tests of Laws IV and V, both under test
The effect of blinding on AI alignment evaluation: Do AI judges' alignment scores keep their direction when the scoring protocol changes?Paper IV.d · Experimental paperScoring controls for the theory’s judged tests
The Stewardship Gene: Do ethical reasoning loops in a system prompt make AI consider those affected more fully?Paper V · Experimental paperEden Protocol: prompt-response study, not a test of the laws
The Load-Bearing Test: Is AI safety a tax on capability, or part of what holds it up?Paper VIII · Experimental paperInforms, without deciding Law V, a law under test
Convergent Evidence for Recursive Amplification: Are separate fields converging on the same principle of recursive amplification?Paper XI · Review paperReviews evidence bearing on Law I, a law under test
Public Benchmark Rescoring: Does blinding change AI evaluation scores on a public benchmark?Paper XII · Methods paperProtocol for testing the reach of the scoring-protocol finding
The ARC Alignment-Scaling Working Report: As AI gets smarter, does it also get more ethical?The ARC Alignment-Scaling Working Report · Working reportWorking account of the alignment-measurement programme

Engineering 3

The Honey Architecture: Does building safety into what a self-modifying system optimises help it stay safe?Paper VI · Architecture paperEden Protocol: candidate architecture, informed by Laws IV and V, both under test
The Eden Vision: What should the intelligence we build be for?Philosophical statementEden Protocol: purpose and values
Eden Engineering: Can correction built into each round keep a self-improving AI good and useful if oversight fails?Engineering Specification · SpecificationEden Protocol: engineering response to Laws II, III, IV and V, each under test

Governance 2

The HARI Treaty (Hardware-Aligned Recursive Intelligence): How could nations govern self-improving AI while the evidence is still coming in?Treaty paper · Working draftGovernance proposal informed by Laws IV and V, both under test
The HARI Treaty (Hardware-Aligned Recursive Intelligence): A proposed treaty to delay the loss of human control over self-improving AIDraft Instruments · Draft treaty text · Working draftDraft governance instrument informed by Laws IV and V, both under test

Overviews and records 6

Synthesis and Roadmap: Which of the programme's claims survived their tests, and which must be tested next?Paper IX · Synthesis paperReviews the programme’s evidence and open tests
Executive Summary: How can self-improving AI remain useful and aligned?Executive Summary · SummaryOverview of the programme and its evidence
Master Table of Contents: Where to start, and how every part of the programme fits togetherMaster Table of Contents · ContentsReading order and document map
The ARC Theory: Does control survive recursive self-improvement?Two-Page Form · SummaryConcise statement of Laws I, II, III, IV and V, each under test
The Dated Prediction Register: When was each prediction made, and did it precede the evidence?RegisterDated predictions and their source records
The Priority IndexRegister · Working draftDated sources and first-publication records

Other fields 3

Polymathy and Neurodivergent Cognition: Does visible high performance show that no support is needed?Paper C · Working paperStandalone work in cognition; outside the ARC Theory
HRIH: Where does recursive self-improvement end, and what would that predict about our universe?MonographSpeculative physical extension; no law depends on it
HRIH research programme: Could intelligence create real universes, and what would count as evidence?Working paper · Working draftSeparable cosmological hypothesis; no law depends on it

The full catalogue

The priority record

The originality of the programme is anchored on a small number of dated artefacts. Each artefact carries a date, a hash, and where possible a downloadable primary source, so any reader can check the dates without trusting the author.

8 December 2024, 02:45:18 UTC

Self-emailed manuscript, five attachments, 221,236 words across them, preserved as an ARC-sealed recipient-form record with OpenTimestamps anchored. Gmail Message-ID CAGPsKAnp-DLDBOeBAh4t9XsqJBKdoe07uORGF86meoSMFnin7w. SHA-256 prefix f0d1f38f. First dated statement of the embedded-correction alignment thesis and the equation in both forms, with all three terms defined and the complement-to-Einstein framing on the record.

/priority-claims

30 April 2025

562-page self-emailed book manuscript, approximately 140,406 words. SHA-256 prefix 09f5b5e1. Expanded statement of the Eden Protocol.

/research/blog/timeline-april-2025-eden-protocol-named.html

2 January 2026 and 6 January 2026

Publication of Infinite Architects. Paperback ISBN 978-1806056200 on 2 January 2026; ebook edition ISBN 979-8302196163 on 6 January 2026. First book-length articulation of the whole programme.

Amazon UK · Press kit

17 January 2026

OSF project created and Paper I released. DOI 10.17605/OSF.IO/6C5XB. First formal derivation of U equals I times R to the alpha and the first preliminary alpha estimates from published data.

DOI 10.17605/OSF.IO/6C5XB · Paper I

Full priority chain (human-readable): /research/priority/. Machine-readable priority register: /research/priority.json. Public priority and claims ledger (41 individually dated entries): /priority-claims.

Concept index and claims register

The programme names, defines and dates its concepts on a per-concept basis. The concept index and the claims ledger are the two entry points readers most often want:

Concept atlas. Named concepts across the programme, cross-linked to their first appearance and to the papers that discuss them.
Priority and claims ledger. Forty-one individually dated claim entries, each with first date, evidence pointers and honest risk notes.
Intellectual priority record. Human-readable priority chain with per-artefact evidence.
Start here. Five-minute map of the whole programme for first-time readers.
Predictions register. A holding page: the twelve early records are withdrawn; registered predictions now live in the registered programme.
Methods notes. Truth-gate, anti-sycophancy, honey dashboard and ARC-Align test protocol.
Proof room. Working proofs and formal derivations.

Contact and citation guidance

Preferred citation for the programme:

Eastwood, M. D. (2026). The ARC Principle: Formalisation and preliminary validation of recursive capability scaling. Paper I of the ARC / Eden research programme. OSF, DOI 10.17605/OSF.IO/6C5XB. https://doi.org/10.17605/OSF.IO/6C5XB

Preferred citation for the book:

Eastwood, M. D. (2026). Infinite Architects: Intelligence, Recursion, and the Creation of Everything. Independently published. ISBN 978-1806056200.

Preferred citation for the priority record:

Eastwood, M. D. (2026). Intellectual priority record for the ARC / Eden research programme. https://www.michaeldariuseastwood.com/research/priority/. Machine-readable: https://www.michaeldariuseastwood.com/research/priority.json

Author identifiers: ORCID 0009-0004-3222-7442. OSF DOI 10.17605/OSF.IO/6C5XB. Wikidata Q140775170. The canonical Person URL is https://www.michaeldariuseastwood.com/; the canonical author page is this page, https://www.michaeldariuseastwood.com/michael-darius-eastwood/. Media, review and research correspondence go via the press kit page.

House rule on citation. Whenever a claim on this site cites the Eden Protocol, it is cited as "the Eden Protocol, named in the 8 December 2024 record and expanded in the April 2025 manuscript". Whenever it cites the equation, it is cited with its artefact (V2 lines 6 to 9 of the 8 December 2024 manuscript; Papers I and II for the general R to the alpha form; the book for the featured R squared presentation). Any single-clause paraphrase that flattens those distinctions is a violation of the site's truth gate and should be reported as an error.