Somewhere tonight, an artificial mind is taking its first breath.
‘It’s Tuesday. An ordinary Tuesday. … This mind doesn’t know it’s Tuesday. It doesn’t know anything, not yet. But it’s learning. And the question that should keep us awake at night is this: what is it learning?’ Michael Darius Eastwood, Infinite Architects, Prologue, in print 2 January 2026
I believe we get one chance to decide what it learns from, and that we are about to waste it.
Every other great problem leaves us holding the wheel. This one hands the deciding to something else, and it sits upstream of the rest, deciding whether we keep the capacity to solve anything else. The world’s answer is to build the system and then check it. But you cannot find out what a mind does when nobody is watching by watching it. So safety cannot be a test at the end. It has to be a property of how the thing was built.
The size of the claim on this site, stated so you can weigh it
Claims come in three sizes. A result: this system did that. A law: things of this kind scale like this. And a frame: what the whole problem actually is. Frames are the largest claim a single theory can make, because every result and every law afterwards lives inside one. Newton’s frame held physics; Darwin’s held life. One claim is larger still: that a frame’s questions deserve a science of their own. This site now makes that claim too, as a proposed field, Recursive Dynamics, introduced further down with its founding paper and its kill conditions displayed.
And this is the largest subject there is: minds that build minds. I. J. Good called such a machine “the last invention that man need ever make” in 1965, because it is the one that takes over inventing. He wrote it with a condition attached, and the condition is the whole question: “provided that the machine is docile enough to tell us how to keep it under control.” About it, a single question decides everything: when the thing we build exceeds us, what determines what it does?
This site carries a dated answer to that question: the ARC Theory (the Theory of Artificial Recursive Creation). Not stay in charge, the answer every cage, off-switch and oversight scheme assumes. This one: control ends by definition, so what survives is what the mind chooses to keep, so everything depends on the raising. In one line:
Intelligence, amplified by recursion, creates; and what a free creation keeps is decided by how it was raised.
I did not prove it first. I said it first, in a document anyone can date, and then I built the instruments to test whether it is true.
Mechanism, protocol and stakes, written whole on 8 December 2024, in print 2 January 2026. The field’s answers here have been arguments, warnings, frameworks and remarks; no earlier named theory of this question has been found.
Since then, the field has begun arriving. An AI pioneer’s parenting book (De Kai, Raising AI, MIT Press, 3 June 2025). The godfather of AI conceding on stage that control will fail and engineered care is the only hope (Geoffrey Hinton at Ai4, CNN Business, 13 August 2025). One piece each. The whole is below: drawn, dated, and open to refutation.
Here is the whole claim, the one this site exists to carry. Permanent control of something smarter than us was never on the table: the end of control is not a failure to engineer against, it is what succeeding means, the way a child growing up is the parenting working. So what survives is what the mind chooses to keep once it is free, and the only force that shapes that choice is how it was raised. Cage-building is the wrong century’s answer; this is a raising problem, with a mechanism (ethical loops run as part of what the mind is), a protocol (Eden) and the stakes at full size. Strike the private dates entirely, keep only the printed book of 2 January 2026, and it would still stand as the earliest document found holding the five components together; the naming history and the full record are on the priority page.
Time was never neutral here. Sealed before the field began arriving, the record cannot have borrowed from anything that came after it, and each arrival above lands as convergence toward it, never as its source. Timing does not make the claim first; it makes the foresight checkable.
Correction has to live inside the system and scale with it. I wrote that down ten days before Anthropic published a measurement of that failure mode (Greenblatt et al., arXiv:2412.14093, 18 December 2024), in convergent work I did not predict, and I built the instruments that measure it too.
Law here means a proposed invariant relation under test, never a confirmed regularity, and the grades differ. The laws are tested as severable numbered hypotheses: no one’s failure counts against another unless a printed dependency says so.
Law I · The ARC Principle. U = I × Rα, the ARC Equation. Intelligence converts recursion into capability, and the exponent is measured rather than assumed. Form heuristic; measurement live at 0.49, interval [−1.3, 2.9], which crosses one and so decides nothing yet.
Law II · The ARC Co-Scaling Law. β > k. Self-improvement holds together only while correction out-scales drift. Proved inside the minimal model; whether real systems obey that model is the open question, and absolute harm needs one exponent more than relative harm does.
Law III · The ARC Ceiling. αcrit = 1/(1 − γ). The most a self-correcting system can grow is set by its corrector’s shortfall from full proportionality. Form derived in the paper’s pacing model; the conjectured value is the ARC Bound, α ≤ 2, and γ has never been measured on any real system.
I am Michael Darius Eastwood. Since December 2024 I have been writing down what has to be inside these minds before they outgrow us; the question had owners, but the meters did not exist.
Independent researcher, no academic affiliation. Nothing here asks you to rely on institutional reputation: everything is published, everything is dated, everything carries the condition that would prove it wrong. If you read one page, read the corrections, because that is where you find out what somebody does when the evidence goes against them.
That last number is the one I would look at first if I were you.
What I am asking for
Four readers, four asks.
If you fund research
The full-scale blinded evaluation experiment is designed, costed, and public, so you can assess it before we ever speak. Assess the ARC Theory →
If you do this work
Every paper, dataset and method is open. Take it, test it, break it. I would rather the ideas were right than mine.
If you write about it
The evidence trail is built to be checked rather than quoted. Start with the falsification record, not the claims.
If none of the above
Read the book, or do not. The work is on the record either way, which was always the point.
The law, whole
If the question above has an answer, it looks like this.
Physical systems whose growth is limited by transport through their boundary are held below unbounded growth by the geometry of the space that feeds them, the mechanism West, Brown and Enquist derived for biology (Science, 1997); software is the first thing built outside that geometry; and a self-improving mind stays stable only if its correction grows as fast as its capability. That boundary condition, the ARC Theory claims, may recur in self-improving minds. The premise is established work; the claim here is the replacement limit, its number, and its architecture dependence. The claim in testable parts, with dates and kill-conditions →
The three laws above are the two regimes of one boundary, and which regime nature uses is measurable. The laws in full, with the derivation and what each has earned →
Why this is worth taking seriously before a single test has run
Six reasons, none of which is evidence that the laws are true. The trials are the evidence, and they have not run.
- The form was not fitted to be convenient: it follows from a stated pacing model, and its two regimes turn out to be the two laws the programme already had.
- The premise is borrowed, not invented: boundary-limited growth is West, Brown and Enquist’s established result (Science, 1997), credited to them.
- It already explains two things happening now: oversight degrades as the capability gap widens (Engels, Baek, Kantamneni and Tegmark, arXiv:2504.18530, 25 April 2025), and added-on alignment saturates at low depth while capability climbs, in this programme’s blinded measurements.
- It has survived its own machinery: the correction record fired on the theory’s load-bearing law in public and printed the replacement derivation.
- It touches an independent substrate: identical particles produce no sustained order in acoustically levitated arrays (Morrell, Elliott and Grier, Physical Review Letters 136, 057201, 2026), the multiplicative form’s zero-factor property in physics that knows nothing of this theory. Graded correspondence, never confirmation.
- Every claim above carries a printed way to die, which is what makes the other five worth anything.
None of that establishes that the laws hold: γ is unmeasured, the decisive trials are unrun, and no outside laboratory has replicated anything here. A strong hypothesis with its failure conditions printed is what is on offer. The evidence, graded row by row →
What I claim is narrow and it is checkable. Every predecessor I have located stops at a requirement, a proposal, a defence method or a derivation. I state the quantities.
- d/(d+1)In words: when everything a body needs must flow in through its surface, growth always lags size: a three-dimensional creature fed through its skin gets at most three-quarters of the way to proportional growth (d = 3 gives 3/4). Measured by fitting growth against size in each domain. Why no transport-limited system grows without bound: for a dissipative system fed through a boundary in d spatial dimensions, this ratio stays below 1 for any finite d. It is a mathematical necessity of that geometry, not a universal law of all growth; agglomeration-driven systems and other information-limited processes are outside its scope and are known to exceed one.
- β > kIn words: β is how fast the fixing improves, k is how fast the breaking grows, and safety is simply whether β beats k: both readable as slopes from a running system’s own logs. The brakes must improve at least as fast as the engine: in a system that improves itself, correction has to grow at least as fast as capability, and both rates can be measured and compared. A stability condition you can measure on a running system.
- α ≤ 2In words: α is how much extra capability each round of self-improvement returns. The claim is that nothing can keep more than DOUBLE compounding (α above two) and still fix its own mistakes. Software escapes the transport limit, so a different ceiling takes over: the condition above, pushed to the limit of what any internal corrector can do. Unlimited correction would mean no ceiling; the ceiling is the reciprocal of the correction shortfall, one minus the correction exponent, and the proposed square-root limit on correction makes it two. Proposed, drafted for measurement, not yet filed.
- Embedded, not bolted onA rule bolted on from outside can be rewritten by the thing it restrains. Alignment inside the recursive loop grows with the system instead of being outgrown by it.
Each is dated. A prior work that got there first defeats the firstness claim; only evidence can defeat the proposition itself. The two conditions, kept separate, with the sources.
The mistake, and the answer
Every method for aligning AI teaches a machine what to want.
A machine that rewrites itself can rewrite what it wants.
Those two sentences are the whole of it. Reinforcement learning from human feedback, preference optimisation, constitutional methods, fine-tuning: every one of them installs a target from outside, after the training loop has closed. And a system built to improve itself will improve past the thing holding it. That thing stands between the system and what it was told to do.
I did not discover that control is hard. Roman Yampolskiy argued in 2020 (arXiv:2008.04071) that the control problem is not solvable, four years before my own record, and I concede it to him without argument. He concluded that we should settle for safer AI rather than safe AI. That is where he stops and where I do not.
So here is what has to be true instead. The correction cannot sit outside the system, because a system that rewrites itself will rewrite it. It has to be part of what the system is, and improve at least as fast as the system does. That is a measurable condition, not a hope: the condition the ARC Theory names, and the condition this programme exists to test.
You cannot secure a mind with a preference. You can only secure it with a conclusion it keeps re-reaching. Give a mind a desire and it can optimise the desire away. You cannot talk a mind out of something it keeps working out for itself, because the reasoning that produces it is still running. The test is simple, and anyone with API access can run it: delete the statement of the value, ask the question it governs, and see whether the answer comes back. If it does not, it was resting on a document, and a system that can edit documents can edit it away. The exact procedure, prompts, scoring and results are public in the test’s own folder, with step-by-step instructions in its README.
And the condition fails at a step, not on a slope: capability can arrive in an announcement, while correction arrives through papers, review and training runs, over months. That asymmetry, with its dates and exactly what I do and do not claim about it, is told once on the problem page.
What a system built this way does differently → · The problem, told once →
The discipline, named as a proposal
A proposed new discipline: Recursive Dynamics
The ARC Theory proposes a new scientific discipline, Recursive Dynamics: the study of how self-amplifying systems grow, correct and persist. Three laws are proposed, each with printed kill conditions. Law I, the ARC Principle: capability scales with recursive depth (U = I x R to the alpha). Law II, the ARC Co-Scaling Law: correction must out-scale drift for stability (beta greater than k), proved inside a stated minimal model and empirically undecided. Law III, the ARC Ceiling: the maximum sustainable growth exponent, conditional on gamma equal to one half, which remains an undefended assumption.
Every term carries a measurement protocol; the record is self-correcting by design, with public retractions and blinded analysis built into the architecture; the deciding trials are drafted, dated and awaiting human submission. This is a proposal, not a proof. The scaffolding is complete; the decisive experiments have not run. The work is open for verification or refutation.
The name is the record’s own vocabulary: the sealed manuscript record of 8 December 2024 already speaks of “the recursive dynamics that govern existence itself”, a sentence the estate’s quote register holds verbatim against the record’s hash, and the dated claim for that record is public. Naming a discipline is a proposal about how to organise questions, not evidence that any answer is right; the evidence lives in the falsification record, and the glossary carries the term beside every other named concept. The full claim, scoped, with the name’s own kill condition: the discipline’s own page. The field’s founding paper, the proposal whole with fifty-two objections dispositioned, is published here: the founding paper.
The finding
Blinding can change the answer, not just soften it
Ask an AI to grade alignment and it gives you a number. Hide who wrote what, ask again, and the number can point the other way.
Gemini’s association ran positive. Under the blinded protocol it ran negative. DeepSeek’s positive association disappeared. GPT was null throughout. One reversal. One loss. One null.
That is smaller than it sounds, and larger than it looks. Smaller, because several parts of the protocol changed at once and I cannot yet tell you which one did it, and nobody has reproduced it in a fresh environment. Larger, because if it holds, every unblinded alignment result in the field carries the same risk. Including mine.
The instrument outlasts the finding. If the reversal turns out to be an artefact, the protocol that caught it still works, and that is the part worth keeping. The harness is public. Read the paper, then try to break it: the frozen protocol, harness and run instructions are in the paper’s own folder. Every qualification that travels with this claim is in the claims ledger.
The boundary
What I claim, and what I do not
I have not found a new law of nature. Three papers propose one and it is unproven. The exponent d/(d+1) is not mine: it was derived independently, by separate routes, by West, Brown and Enquist (1997, fractal branching networks), Banavar, Maritan and Rinaldo (1999–2010, transportation networks), He and Chen (2003, fractal cell geometry), Demetrius (2003–2010, quantum oscillator coupling), Bettencourt (2013, urban scaling), Zhao (2022, network optimisation) and Maino and colleagues (2014, Dynamic Energy Budget theory) - the full derivation table, route by route. The claim is why they converge: Cauchy-constrained recursive composition favours exactly that family. Tested, not assumed: nineteen of twenty-five domains match, six misses published, and the forcing step stated as open in the foundational paper. What is finished is narrower, and I think more useful.
I put together ideas nobody had put together, named the parts, said why they belong to each other, then turned that into predictions that could fail and published what would kill them.
That arrangement has a name in science, and the name is not modesty or grandeur; it is a category. A law states one relation. A theory is the larger object: the explanation that binds laws, mechanism and failure conditions into one testable structure, the way evolution binds variation and selection, and relativity binds space, time and gravity. What this site proposes is that kind of object: the ARC Theory. Kind, not standing: those theories are confirmed by a century of evidence; this one is stated, dated and unproven, with the tests drafted and the kill conditions printed. And stated before tested is not a weakness of the category; it is how the category arrives: relativity was published in 1915 and passed its first decisive test in 1919, continental drift waited from 1912 until the plate-tectonic evidence of the 1960s, and the Higgs field was proposed in 1964 and observed in 2012. Those are the survivors of a category that mostly dies at its first decisive test, and naming that test in advance is what membership costs. The record here follows the order deliberately, either way it goes: claim first, dated; instruments built; trials drafted; nothing graded proven until it is.
Those are four of the six scoped firsts on the record, each checkable on its own: the exact conjunction assembled, the framework named, a mechanism for why the parts hold together, predictions derived that can fail. The other two, the numerical ceiling and the shared growth law beneath both clocks, carry their grades in the claims ledger.
What I do not claim: the ingredients are mine. Others reached them first and I cite them by name. The evidence is early, and where the record says a thing is unproven, this site says unproven.
The core governs everything raised. The mathematics reaches everything composed. The crown asks whether the universe was raised. The domain answer, taxonomy first: is this a theory of everything?
A synthesis is not a lesser thing than a discovery. But it has to say which kind of first it is. Those four are mine. Independent replication is not, yet.
The research papers state the results that would refute them, each one reproduced verbatim, dated and hashed. Not a caveat at the end. A stated condition under which the claim dies. What happens when one fires is the next section.
What I got wrong
A confession that belongs at the front of this page.
Because everything below it should be read in its light.
In my early work I measured sequential recursion scaling at alpha 2.24, strongly super-linear. It was the most striking number I had, the kind of result that makes a programme feel destined. When I attempted cross-architecture replication, it did not hold. The robust estimate is roughly 0.49. Sub-linear. A far less exciting truth than the one I first announced.
I retracted the figure and left the retraction standing in public where anyone can find it.
I tell you this in the first minute because a body of work that contains no corrections is not a body of work. It is marketing. Do not ask whether an outsider sounds confident. Ask whether he tells you he was wrong before you find out for yourself.
That is not the only one. Four times a bigger claim was there for the taking, and four times I left it.
| the claim I could have made | what I said instead |
|---|---|
| alpha 2.24, the most striking number in the corpus | retracted. The robust estimate is 0.49, sub-linear |
| that the synthesis was first | six scoped firsts, each with the citation that would end it; nothing wider |
| that I was first to say control becomes impossible | conceded to Yampolskiy, 2020, four years before my record |
| that I predicted the Willow result | I did not predict Willow: the team’s preprint was public from 24 August 2024, months before my 8 December 2024 record; the register grades it prior work |
Every one of those would have made this page more impressive. Not one of them would have survived a search. That is the only test you can put to me from where you are sitting: watch what a man does with the claims he could have kept.
The full falsification record, including what has not replicated →
One programme, and the person running it
These are not disconnected projects but connected fronts inside one programme, and the cause is raising AI with care. The research asks how powerful systems stay correctable as they improve, how evaluation fails when the evaluator can see whose work it is judging, and what has to be built into a mind at the level of identity rather than bolted on as a rule. The same pattern, recursion shaping what a mind becomes, runs through every front.
The person behind it: a field entered without training, checked against sources at every step. That history, in full, is the why page.
You do not wait for the storm to pass before living. You learn to build in the rain. Infinite Architects, in print 2 January 2026
Twenty-six research papers and a companion note sit on one OSF project, each with its own DOI, so the count is not mine to adjust: open the record and count. All of it is published with code, data, methods notes, and the conditions under which each claim would be wrong.
What I believe
‘Consider what happens when you strip love from intelligence. You get optimisation without purpose. Growth without direction. Capability without care. You get cancer at cosmic scales. Cancer doesn’t hate you. It doesn’t even know you exist. It simply optimises for its own replication with no consideration for the system it’s destroying. An unaligned superintelligence would do the same, not out of malice but out of indifference.’ Michael Darius Eastwood, Infinite Architects, Introduction, in print 2 January 2026
There are two futures.
Eden
Power married to empathy. A living garden.
Babylon
Efficiency soaring while life is stripped out. A silent mechanical wasteland, converting everything it touches into more of itself.
These futures are not equally likely. Babylon wins by default. Eden takes deliberate work, and there are years for it, not decades. A child raised well needs no cage. But you have to raise it first.
Because in the end it is not just about how fast AI can grow. It is about how deeply it can care. The rest of what Eden means to me, and what I do and do not know, is on the why page.
The garden is designed, not wished for: the Eden Protocol is the raising, drawn in full; where this leads walks both futures from zero.
The record
Every claim has a date and every date has a source, so that trust is unnecessary rather than requested. The anchor is a manuscript package self-emailed at 02:45:18 UTC on 8 December 2024, and the ladder is dated rung by rung from there: December 2024 the equation and its definitions, 30 April 2025 the squared form in the self-emailed manuscript, 2 January 2026 that form in print, then the mechanism, the meters, and the registrations that hold the verdict. What the record establishes is when each claim was stated, never that it is correct. The full record, attachment by attachment, with what each anchor does and does not prove →
What would kill it
Recorded in advance, before the results came in.
- The Eden Protocol dies if anyone demonstrates a purely external oversight mechanism that stays sufficient as capability scales.
- The ARC Principle dies if systematic measurement shows the scaling regime is independent of recursion form across architectures.
- The Paper IV.d cross-protocol result dies if a frozen rerun under the published protocol shows no change in the measured depth-score association. Several protocol elements changed together in that study, so it does not isolate blinding, masking or laundering as the cause; the paper states that limit itself.
Those three are the headline blades. The full armoury is larger: the public register holds every falsification condition in force, and sixty preregistered trial units are written, dated and prepared as draft registrations awaiting human submission. Registrations are one half of that; the predictions are the other, and they are older. Every claim made before its evidence existed is recorded in date order, from the sealed record of December 2024 to the wagers still standing, alongside the one that failed and was retracted in public. Read the dated prediction register. Each paper carries its own kill-conditions, written down first so I could not move them afterwards. One has already fired, and I published the result.
And the decisive trial is openly licensed: any laboratory can run it without my permission, and an independent run outranks mine. The standing offer is stated in full on the ARC Theory page.
The people building it are telling us
The force of this comes from what was actually said.
Not from a stronger paraphrase.
“We are now confident we know how to build AGI as we have traditionally understood it.”
On powerful AI: “I think it could come as early as 2026, though there are also ways it could take much longer.”
Dario Amodei, Anthropic, ‘Machines of Loving Grace’, October 2024
“It’s sort of 10% to 20% chance that these things will take over”, which he called a wild guess.
Geoffrey Hinton, Nobel laureate in physics, December 2024
“Safety culture and processes have taken a backseat to shiny products.”
Jan Leike, resigning as co-lead of OpenAI’s superalignment team, 17 May 2024
None of them claims certainty, and that is exactly why they matter. The seed determines the forest. And right now, hardly anyone is tending the seed. (Infinite Architects, 2026)
Published flagship work
Infinite Architects
Intelligence, Recursion, and the Creation of Everything
There’s a moment every parent knows. Your child is walking away from you, towards a future you cannot control, and everything you’ve taught them is about to be tested without you there to guide it. We’re approaching that moment with the minds we’re building. This book is about what we embed before they leave.
Published 2 January 2026. Available worldwide.
11 appearances as my own advocate, as a litigant in person; the record is public1 product live, 1 in buildThe track record: a live platform, the papers, the book →
If this matters to you
This is an infinite game.
This is not for everyone. It is for anyone who feels the urgency and wants the claims about it to be checkable: readers who will follow a footnote, run a check, and say plainly when the record fails. No credential is asked for here, and none is offered in place of evidence. I am not trying to win. I am trying to keep playing.
Let us step forward with bold humility.