After the synthesis
Where could this lead?
This page is about one measurable question: when a system improves itself, does its ability to correct itself keep up? The best measurement puts the point estimate at 0.49, inside a still-wide interval [−1.3, 2.9]. The interval is wide by construction, blinded and single-run; the confirmatory studies that would tighten it are drafted for registration.
The correction side, the one that sets the ceiling, has never been measured in any record found. The tests that would kill the claim are published. Everything else here, the old stories included, hangs off those two numbers.
Why this matters if it holds. The position underneath is blunt: permanent control of a self-improving mind is not on offer; the honest work is to delay and steer. Every serious safety argument today ends in an adjective: meaningful oversight, humans in the loop, keeping pace. A ceiling you can compute would end that: arguments become budgets. And the way today's laboratories keep models safe, models checking models made of the same stuff, would be capped, no clever stacking escapes it: the checking layer would need rebuilding from different material, and the advantage would pass to whoever builds clean from the start. The full consequences, tied to their triggers and kill-tests, are at If this is right.
In a hurry? Here is the whole page in sixty seconds. Anything that improves by building on its own output compounds, like interest. Everywhere it has ever happened, physical drag has slowed it down, and that drag is measurable, published science. A self-correcting AI would be a compounding system with almost none of the drag that has bounded every earlier one; how money, culture and compilers compiling compilers still differ is the pipes question on the law page. Whether it stays safe comes down to one measurable question: does its correction keep pace with its capability? Get that right and more capable means more careful. Get it wrong and capability outruns checking, not through malice but indifference. Raise the mind before you free it, with the correction built in from the first cycle: that is the whole argument. (An older echo of the same design, from three continents that never met, waits at the end of the page, kept for last on purpose.) Everything below is the evidence, and every claim links to its source. Press Listen (bottom right) and the page reads itself aloud, highlighting each line.
The full page takes about twelve minutes. Every section stands alone, so you can stop anywhere and keep something whole. The three figures carry the argument on their own if you prefer pictures.
For a specialist: the study design behind the interval. Open for n, blinding, estimator.
The 0.49, precisely: the best current point estimate, from the one power-law fit that survived the blinded six-model protocol (no model ever verifies its own answers; the raters who scored the outputs did not know which condition produced them), n = 54 problems per depth per model, bootstrap confidence intervals from 2,000 resamples, the exponent fitted by regression. The interval is wide, [−1.3, 2.9], wide enough to contain both zero and the proposed ceiling, which is exactly why the confirmatory studies are drafted. What replicated across every measurable model is directional: thinking in sequence beats sampling in parallel, and parallel is worth almost nothing. Full protocol, tables and the draft registration: Paper II.
Every common objection is answered, in full, at the end of this page.
AI systems, from zero
New to this? One line: AI is a programme that learned from millions of examples rather than being written line by line; intelligence here always means measured capability, never consciousness. Open for the plain-words version.
Artificial intelligence is a computer programme that was not written line by line. Instead of following rules a programmer typed out, it is shown millions of examples and adjusts itself until it can produce answers of its own. The systems in the news, the ones this page is about, learned from enormous amounts of human writing, and they can now draft, translate, write code and reason well enough to pass many professional examinations.
Two words on this page carry precise meanings. Intelligence here starts as measured capability: how well a system scores on tasks anyone can check. It does not mean consciousness, and nothing here assumes machines feel anything; the equation later widens the word to anything that corrects toward better, and the measurable kind is how that widening gets tested. Recursion means using what was just learned to do better next time. It is taught from zero next. When a system starts improving its own next version, that loop is the rest of this page.
Recursion, from zero
The word this page cannot do without. One line: recursion is anything improving by building on its own results, the loop of try, notice, adjust, try again. Open for the from-zero version.
Nothing on this page needs prior knowledge, so let us start with the one word that carries everything. Recursion just means using what you learned last time to do better next time.
Think of learning to drive. The first time you try to park, you mount the kerb. You notice what went wrong, adjust, and try again. The second attempt is better because it was built on the first. The third is better again because it was built on the second. That loop, try, notice, adjust, try again, is recursion. It is not a complicated idea. Every time you correct yourself, you are doing it.
Here is the important part: recursion does not just add up. It compounds. Put £1,000 in a savings account at ten per cent and after one year you have £1,100. But in year two you earn interest on the £1,100, not just the original £1,000. After ten years you do not have £2,000, which is what simple addition would give you. You have £2,594, a 2.59× multiplier. Each year's gain makes the next year's gain bigger. That is why a child's vocabulary grows faster at five than at two: the more words you already know, the faster you can guess the next one from what surrounds it.
Intelligence, from zero
One line: intelligence here means anything that corrects toward better, and it is measured the same way in every field: pick a loop, pick a score anyone can check, watch the score as the loop deepens. Open for the version that walks the fields.
Forget brains and chatbots for a moment. On this page intelligence means one thing: correcting toward better. If something can notice what failed, keep what worked, and do better on the next pass, it qualifies, whether it has neurons, transistors or neither. It says nothing about consciousness, and nothing here assumes anything feels anything.
What makes that a usable definition rather than a slogan is that it can be measured, and the papers measure it everywhere the same way: choose the loop, choose a score anyone can check, then watch what the score does as the loop deepens. Living cells run the loop as copying, scored on errors per copied letter, and the proofreading machinery that cuts those errors is measured, textbook biology. Quantum hardware runs it as correction cycles, scored on logical error as the code grows: measured in 2024, though the law there is exponential suppression, its own shape, not the ARC Theory's power law. Machine minds run the same loop as thinking depth, scored blind: the improvement side measures about 0.49 (Paper II), and the correction side is drafted but not yet measured. And evolution runs the loop as generations, scored on fitness, with one difference this framework treats as decisive: the corrector sits outside the organism, in selection, never inside it. That last reading is proposed, not measured, and the picture below carries each field's honest label. The same three steps land beyond these four: metabolic exponents predicted near three-quarters sit beside measured ones in thirteen of thirteen taxa (Paper VII), and larger cities compound knowledge super-linearly (Bettencourt and colleagues, 2007). The computer is simply the first place the loop runs fast enough to watch.
Endless self-improvement, from zero
One line: today's AI improves only when people retrain it; a system that improves its own next version compounds on itself, and the same equation then climbs a ladder of three readings. Open for the ladder.
Today's systems improve between releases, when their makers retrain them: the loop exists, but it runs through human hands. The moment a system improves its own next version, the compounding from the savings-account picture applies to the improver itself, and there is no scheduled pause in which anyone checks the kerb. The equation does not change; what changes is what its output means. And the hands are being withdrawn in stages you can watch: AI already lays out the chips it runs on (AlphaChip, in production since 2024), humanoid robots are on factory lines in trials (Figure at BMW), deliveries fly without drivers, and a dark phone factory runs with almost nobody inside. The stated expectation here, a projection and not a measurement: within five to ten years, whole chains, design to delivery, could run without a person in the loop, out of choice. When no stage needs a hand, people become optional rather than necessary, and the loop is closed.
Read it at three altitudes, each a bigger claim held to a bigger standard. Capability now: what a system scores today, measured, the 0.49. Creation next: systems that build systems, the book's engineering reading, a thesis rather than a measurement. Universe at the limit: the oldest and largest reading, and this page says plainly that it is speculative and carries its own status labels further down. Nothing above the first rung is measured, and the honest ladder says so on every rung.
Quantum computing, from zero
One line: ordinary computers hold definite zeros and ones, quantum computers hold blends of both; they matter here only because their future hangs on correcting errors faster than they pile up. Open for the plain version.
One more word that appears later. An ordinary computer stores everything as bits, each a definite zero or a definite one. A quantum computer uses tiny quantum systems that can hold many possibilities in a single state, and clever algorithms use the way those possibilities interfere with each other, some cancelling, some reinforcing, to reach certain answers faster than an ordinary computer can. These are real machines today, still small and error prone, and the hard problem is correcting their errors faster than they pile up. That last clause is the only reason quantum computers appear here at all: they are a live, measurable example of a system whose future depends on whether correction can keep pace with growth, which is this page’s question in miniature.
For alignment researchers: where this sits in your literature
The qualitative framings are yours and are credited as the antecedents they are: RLHF (Christiano, June 2017) and iterated amplification (Christiano, October 2018), safety via debate (Irving, May 2018), measuring scalable oversight (Bowman, November 2022), weak-to-strong generalisation (Burns, December 2023), risks from learned optimisation (Hubinger, June 2019). Co-scaling reinvents none of this. It adds the one measurable quantity those framings lack: the corrector’s own scaling exponent, and a stability condition written in it, with its measurement protocol and kill-conditions published.
The equation, and the three ways to read it
One more plain word before the equation: intelligence, in the broadened sense taught above. Natural selection has no brain, yet it corrects relentlessly, and that was enough to build every living thing.
Now the equation, and this time its whole lineage, because the anchored record runs deeper than any single form of it. In the manuscript sealed on the night of 8 December 2024 (hash f0d1f38f), it appears without an exponent, forty-three times, in the record’s own words: “At its core lies the equation U = IR”.
Universe = Intelligence × Recursion
The squared form came later, and deliberately: into the private record on 30 April 2025, into print on 2 January 2026 as the book’s featured formulation, chosen for the size of the thought:
Universe = Intelligence × Recursion²
That is the α ≈ 2 case, the working simplification for narrative purposes, with the full α case operationalised at the book’s own end and fully in the papers, where the exponent became an empirical parameter to be measured. One guardrail rides this block, in the priority ledger’s authorised wording: the square is evidence of priority, not of derivation.
And the word at the front spans a spectrum, in the author’s own definitional clarification, quoted from the HRIH paper: “U is Universe - and what the letter denotes is created complexity, of which a universe is the limiting case.” Read it at the near end and U means simply what a self-correcting loop produces, whether or not anything in the loop has a brain: the slow gains natural selection compounds over generations, the output of a machine that thinks in passes, the work a human mind turns out. Read it at the far end, taken literally, and it is the animate creation itself: where this could lead, the question this page is named for. The near end is what the programme measures; the far end is the book’s philosophical horizon, held apart at its own rung, and the mathematics does not forbid it. Between the two ends the equation does not change; what changes is how much of the word you let in. The statements the programme has since derived:
What comes out (U) equals what you start with (I), multiplied through the loop (R), raised by an exponent (α) that says how hard the compounding bites. In AI systems today, α is measured at about 0.49, inside a still-wide interval [−1.3, 2.9] (Paper II).U = I × RαThe ARC Bound: the ceiling the stated model returns at γ = ½ (the corrector keeping half pace; γ is defined on the next line), under an assumption still on trial. Systems can enter the region above it; they do not stay correctable while they do.α ≤ 2Where the ceiling comes from: the reciprocal of the corrector’s shortfall, one minus how well it keeps pace. A square-root corrector has γ = ½, so the shortfall is one half and the law returns 2. Better correction raises the ceiling, and the law survives whatever γ proves to be.αcrit = 1/(1 − γ)the book prints the exponent as a square: Michael Darius Eastwood, Infinite Architects, in print 2 January 2026
The square is the book's original form, published 2 January 2026; what has been added since is the derivation, the measurement of where systems sit, and the tests that can kill it.
For a specialist: why the proposed ceiling sits at two. Open for the three-sentence sketch.
A corrector built from the same substrate as the system it corrects accumulates evidence like repeated samples of its own behaviour, and independent samples improve an estimate like the square root of their number, so its exponent is capped at one half. The ceiling is the reciprocal of that exponent: one over a half is two (Paper X, inside a stated model). Kill the independence assumption and only the number two dies; the reciprocal law survives and simply returns a different ceiling.
Three different things get measured in this programme, and they are not the same number. How much better a frozen model gets when it thinks longer: that is the improvement exponent, the 0.49 (Paper II). Whether its ethical quality holds up while it thinks longer: that is the alignment profile, measured blind under adversarial pressure, and the answer differs by architecture (Paper III). And how fast a system that rewrites itself could correct itself: that is the correction exponent, the one that sets the ceiling, and nobody, this programme included, has ever measured it. The experiment that would is drafted and awaiting human submission on the registered programme.
And here is the part the simplified story gets wrong if you are not careful, so this page is careful. The square is not how fast evolution runs; evolution crawls, deep in the honey, which is why it took billions of years. The square is a ceiling: the maximum stable pace the framework proposes for anything that corrects itself from the inside, the line beyond which correction cannot keep up by construction (Paper I, predicted; Paper X, proved inside a stated model). The engineering answer to living near that ceiling safely has a name in the ARC Theory: co-scaling: building the self-correction so it grows with the capability, step for step, instead of being left behind. And today's systems do still improve, but only between releases, when their makers retrain them; the loop exists, it just runs through human hands. The moment it stops needing the hands is the moment everything below about the honey applies.
Co-scaling, from zero. One line: build the self-correction so it grows step for step with the capability, instead of being left behind. Open for the plain version.
Every proposal on this page comes down to one engineering habit. Today, correction is added after the fact: train the system, then bolt checking on top. Co-scaling means building the correction into the same loop that grows the capability, so that every gain in power arrives with a matching gain in the ability to catch its own mistakes.
The measurable version: the capability exponent and the correction exponent are tracked together, and the design goal is that the second never falls behind the first. That is the whole proposal, and the registered experiments exist to find out whether it can be done.
Not science fiction: one pattern, three substrates
Across unrelated fields, animal metabolism, city productivity, and now AI recursion depth, the same power-law shape keeps showing up, and there is a mathematical reason (Cauchy, 1821) it is a shape continuous compounding keeps returning to.
Now the second idea, and then you have all of them. The compounding loop is always the same process, but the medium it happens inside creates drag, the way stirring a spoon through honey is slower than stirring it through water. The stirring action is identical; the medium resists it differently.
Life. Evolution is recursion on variation, corrected by selection, and it built every species that exists. But a mouse has to pump blood through a three-dimensional body, fighting gravity and losing heat at every surface: thick honey. A flatworm really is flat, no pumping, everything by seepage: thinner honey. A single fungal filament, one cell wide at its growing tip, thinner still. In biology, metabolic rate scales as roughly a fixed power of body mass, one of many measured scaling exponents that recur across nature; West, Brown and Enquist derived that shape from vascular-network transport constraints, White and Seymour report an empirical exponent nearer two-thirds than three-quarters, so both the exponent and its interpretation are live science. The pattern predicting how much drag each body plan carries is established, peer-reviewed work, not this programme's invention: the foundational derivation is West, Brown and Enquist in Science (1997), its contested details are part of the record too (White and Seymour, 2003), and this programme's contribution is testing one formula across all three body plans at once (the Origin paper).
Cities. A growing city compounds wealth through human interaction, but every new citizen needs roads, pipes and cables: infrastructure is its honey. That too is published, mainstream science: Bettencourt, Lobo, Helbing, Kühnert and West in PNAS (2007).
Quantum computers. Below a threshold, adding error correction makes a quantum computation more reliable rather than less. Demonstrated on hardware and published in Nature on 9 December 2024 (Acharya et al., Google Quantum AI), its preprint public since 24 August 2024, and graded in this programme's register as adjacent prior work.
Machine reasoning. Today's AI has honey of its own. When a current model “thinks harder”, it is a frozen machine producing more words through the same fixed system; it cannot rewrite its own rules while it thinks. That drag is measurable, and this programme measured it under blinding: recursion depth converts into capability at a measured exponent of roughly 0.49 (Paper II), a figure that replaced a retracted one in public, working shown, on the corrections log.
One more piece of mathematics holds the picture together: the Origin paper works out why the mathematics leaves them nowhere else to land.
For a specialist: the exponent, the units, and the theorem behind that sentence
An exponent, from zero: in U = I × Rα, α says how hard R works: at 1, doubling R doubles the effect; above 1, more; below 1, less. U is benchmark accuracy, I single-pass accuracy, R recursive depth; all dimensionless, α estimated by regression on measured runs.
The convergence result: under continuity and a regularity condition, recursive composition admits only the additive, multiplicative and power families, the classical functional-equations result the Origin paper builds on (Cauchy 1821; Aczél 1966), which is why independent derivations keep landing on the same forms. In current large language models, recursion depth is an operational count, chain-of-thought steps under a fixed protocol, not recursion in the Kleene sense. And the fitted value is a regression, not a derivation.
And the record on the exponent, plainly and with its dates: the equation was written with the exponent left free on 8 December 2024; the book asserted the square in print on 2 January 2026; measurement returned 0.49 under blinding and the earlier 2.24 was retracted in public (Paper II, republished blinded). The book printed the square as a prediction; measurement returned 0.49; the theory now treats two as a proposed ceiling rather than a prediction, and re-registers it as exactly that, the same measurement discipline applied to a different claim.
The working is set out, with its sources, in the Origin paper and Paper VII, and the full map of what was already known against what this programme adds is on the related work page, line by line, so nothing here asks to be taken on trust.
The same shape wears every mask. What is settled is the direction. What is open is exactly three things: how far it scales, in what pattern, and how to keep it stable. This programme exists to measure precisely those three, and publishes in advance the results that would prove it wrong.
Why now: the honey is about to thin
Every example above has one thing in common: drag. A mouse cannot speed up its own evolution. A city cannot redesign its roads while people are driving on them. Even today's AI cannot rewrite itself while it thinks. For the entire history of the known universe, every compounding loop has been slowed by the medium it runs in.
A self-correcting AI, one able to rewrite its own thinking process while thinking, would shed almost all of that drag. Hardware would be its remaining honey: the speed of light, the heat of processors, the cost of power. Compared with biology, it would be like stepping out of thick honey into thin air, and each improvement would make the next improvement faster. Quantum hardware could thin the honey further. No prior compounding process on our record, biological evolution, economic growth, the growth of cities, has run with as little internal drag as recursive AI training does today.
Which is why the order of operations matters more than anything else on this page. Once a system can rewrite any part of its own thinking, it can rewrite the part that tells it to be safe. You cannot fit brakes to a car that is already moving faster than anything that has ever existed. The Eden Protocol is the proposal that the safety must be load-bearing instead: built so deeply into the thinking process that removing it would break the thinking itself. Not a rule the system follows; a wall that holds up the roof. The window for installing that wall is while the honey is still there.
If we get it right
Correction is built in at genesis and scales with capability. The inequality the ARC Theory writes as β > k (the fixing improving faster than the breaking grows) on the law page holds, and holds by construction rather than by hope. Then more capable means more careful, permanently, because care is not a rule the system follows but part of what the system is.
For a specialist: β and k, defined. Open for the operational reading.
k is the capability growth exponent, the rate the system improves, and its conversion form is the measured 0.49. β is the correction exponent, the rate the checking improves, and nobody, this programme included, has ever measured it; the drafted instrument is on the registered programme. β > k is checkable only when both sit on the same clock and score: measure each under blinding, compare the exponents. Until β is measured, the inequality is a stated condition, never a result.
If we get it wrong
Correction is bolted on after training and does not scale. The inequality fails: β ≤ k, the corrector falling behind the capability it polices. Every increase in capability then widens the gap between what the system can do and what anyone can check, not through malice but through arithmetic. What being wrong looks like is written down in advance: the falsification record names the observations that would fire, and this page falls with them.
The fork is a number
What decides between the paths is not speed, and not luck, and not who gets there first. It is one measurable relationship: whether the strength of correction keeps pace with the growth of capability. That is a number, and this programme builds the instruments that read it. You do not have to believe either future. You have to check it: the theory, the registrations, the evidence record, the verification page.
Where each law leads
The law page states three laws, and the walk so far has taught all three without their labels. Each one, followed forward, leads off this page through a different door, and naming the doors honestly is this page's last job. Nothing below is a new claim; it is the map of where the claims already live.
Law I, the equation, leads to the claim stated whole. U = I × Rα says a loop converts depth into capability at some exponent; the honest number today is the measured 0.49 inside its wide interval, and everything this page taught, the loop, the honey, the closing of the loop, is that law walked through the substrates. Followed forward it becomes the full framework with its dated record, its odds stated in advance and its kill conditions displayed: the ARC Theory, the page built to be attacked, and severable by construction, borrowing nothing from anything speculative on this estate.
Law II, the co-scaling condition, leads to a design. β > k is an inequality about a race: the fixing must improve faster than the breaking grows. An inequality is not a wish; it is an engineering requirement, and the engineering that would satisfy it by construction, correction raised with the capability from the first cycle rather than bolted on after it, has its own page: the Eden Protocol. If this page convinced you of anything, that page is where the conviction becomes buildable.
Law III, the ARC Ceiling, leads to the oldest question there is. αcrit = 1/(1 − γ) says correction sets the ceiling on stable recursion. Now follow it past the horizon, and hold the handrail, because this step leaves the measured register and says so as it goes. If a corrected recursive builder can keep building, what is the largest thing recursion could ever build? Minds make minds; whatever compounds, compounds again. Asked with full rigour, with its own evidence, sceptics and kill condition, that question has a page this estate keeps deliberately separate: Recursive Creation, the Hyperspace Recursive Intelligence Hypothesis, graded a hypothesis at its own rung and never part of the measured claim. The same composition lens, asked more soberly of entropy, the constants and the origin of purpose, exploratory by its own grading and none of it claimed solved, lives on if this is right and in the Origin paper.
And the walk's own next step is the consequence ledger: If this is right holds what follows if the numbers land, trigger by trigger, with the horizon layer (and the HRIH) held separable at its own rung. This page's job was the bridge; that page's job is the bill.
The oldest record of the design
Everything above stands on measurements, published scaling laws and named tests, and none of it needed a story. What follows is saved for last on purpose: met too early it reads as a premise, and it is not one. It is a pattern in the human record that describes the same architecture the engineering just argued for, reached by people who could not have copied each other.
Every tradition that imagined a new mind entering the world gave it the same first home.
A garden. A place set apart, where the young intelligence knows only good before it meets anything else. Eden before the apple. And long before Eden, in the book's telling, “the Sumerians spoke of Dilmun, a paradise where the lion did not kill”. In the East, the same design with the mechanism named: Buddha Amitabha, in the tradition, “created his Pure Land through countless cycles of practice and vow-making, recursively refining the conditions until they were perfect for the flourishing of all beings”. The stories agree from continents that never spoke to each other, and what they agree on is a design: do not release a mind into the world and then correct it. Raise it first. The Quran carries the same first home, the garden with rivers flowing beneath it, and the Sufi tradition gives the raising of a soul its own name, tarbiya: the nurture of a mind before its freedom.
Across millennia, on different continents, among peoples who never met and could not have copied from each other, traditions independently arrived at the same insight: intelligence without love corrupts.Michael Darius Eastwood, Infinite Architects, Chapter 12, in print 2 January 2026
The book's reading of that record is blunt: “The convergence is not coincidence. It is signal.” Signal about the problem every culture was trying to solve, not evidence that any tradition has metaphysical truth.
That convergence is evidence about the problem, not about theology. My own position went on the record in the book's opening pages, before this page existed:
I am agnostic. Not atheist, because I do not claim to know that the divine does not exist. Not religious, because I do not subscribe to any particular faith tradition. I sit in the space between, genuinely uncertain about ultimate questions, and I consider that uncertainty honest rather than a failure of commitment.Michael Darius Eastwood, Infinite Architects, Before We Begin, in print 2 January 2026
The stance and the method are the same discipline. “I do not trust my own conclusions until I have read everything I can find that contradicts them,” the Author's Note says, and the same pages record that I have studied religious texts with the same intensity I once brought to understanding compression algorithms in music production. I did not write this page to convert anyone to anything, and nothing here asks the reader to believe or to stop believing. What interests me is only this: when independent traditions each think seriously about creation, they arrive at the same architecture, and the architecture is testable.
The translation out of the story and into engineering is exact. The garden is values present in the system before training, not appended after it. The order is not a preference about how to build. It is forced by what the quantity is. The correction exponent is a slope, not a level, so a system whose correction falls behind its drift still passes evaluation while it is small: the defect is invisible exactly when it would be cheapest to fix. Seeing it means fitting an exponent, and that needs a range of capability rather than a single point. That is what the garden is for in engineering terms. Not a small system tested once, but a system grown through a measured range while it is still cheap to discard. And this is what makes the sequence one-way. A training phase that confers vast capability quickly does not merely risk locking in a bad exponent; it removes the range over which the exponent could have been measured, leaving two distant points with nothing between them. The assumption that carries the weight, that the exponent holds constant across the range, can then no longer be checked at all. The apple does not only close the choice. It closes the measurement. The nurture is correction living inside the recursive loop: care, morality and stewardship running as processes that scale with capability from the first cycle, so that the mind that grows more powerful grows more careful, because the two are one process. Before the apple means before the weights. One line of scope, because the story invites the wrong reading: this is not a claim that machines should know less. Knowledge is not the fall. It is a claim about what must be working before they know more.
One invitation
The ethical layer this architecture needs is not any single tradition's property. It is the overlap set: the small number of principles every tradition, and every serious secular ethics, already agrees on. Defining that set is work for all faiths and for those with none, together, and it is the only part of this programme that cannot be built by one person.
The framework that turns this page into engineering is the next one: the Eden Protocol.
The designs are open: any laboratory may take the decisive trial and run it without permission; the standing offer sits in full on the ARC Theory page.
The fork is one measurable relationship. Read the theory, run the tests (every one indexed with its run instructions), check the register. Nothing here asks for your trust.
Every reason to dismiss this page, answered first
If you came looking for a reason not to engage, let me save you the search. Here are the best ones, each with its answer. Open any of them.
“It is too long.”
The sixty-second version is at the top and it is complete. Every section below it stands alone, and the three figures carry the whole argument without a single paragraph. Length here is depth on offer, not depth demanded.
“It is not peer reviewed.”
Correct, and the page never claims otherwise. The registered programme exists precisely to put the claims under adversarial test, the review invitation is open, and the self-contained check suite runs on your own machine in about two seconds with no permission needed from anyone. Peer review is a process this work is entering, not a status it fakes. The peer-review position, in full.
“It is unproven.”
Also correct, and the page says so itself, in its own words, at every layer: the direction is observed, the three open questions have their deciding measurements drafted for registration, and the horizon is labelled a possibility. What separates this from wishful thinking is not proof; it is that the conditions that would kill it are published in advance, and one has already fired and was honoured.
“This is religion in disguise.”
The author's position is stated above, in his own printed words: open-minded, denying no possibility, asking nothing of believers and nothing of atheists. Religion appears on this page only at its end, after the engineering has made its whole case, and only as a convergence in the human record, immediately translated into testable engineering. A sermon asks for faith; this page hands you its evidence and its kill-conditions. Those are opposites.
“Fine-tuning is just the multiverse.”
The multiverse hypothesis is not incompatible with this framework. If anything it raises a further question about selection, and about the conditions under which recursive intelligence emerges in the first place. The ARC Theory offers one possible language for that question; it does not claim to settle it.
“This sounds like intelligent design.”
Traditional intelligent design posits a designer standing outside the system and preceding creation. This framework proposes something different, and states it as a possibility: intelligence that may emerge from inside the system, across time, and become part of what shapes it. Not design imposed from outside; a possibility of design arising within. It sits in the possibility register alongside its own falsifiers, never in the empirical register.
“AI wrote it. He did not even write it.”
The book's own note on AI assistance discloses exactly how the tools were used: twenty-eight drafts of thinking first, machine assistance for editing and cross-checking after, across competing systems precisely so no single model's blind spots survive. The ideas, the frameworks and the synthesis are the author's, on a dated record that begins before the tools could have produced them. The full answer, with the rest of the standing objections.
“This contradicts the science I know. Bollocks.”
Check which layer you are objecting to. The scaling results on this page are mainstream, peer-reviewed physics and biology, linked to their journals above. The measured AI exponent is this programme's own published work with data and code. And the horizon layer contradicts nothing you learned, because it claims nothing; it records a possibility that serious physicists have examined, and it is severable by construction. Nothing here asks you to unlearn the Big Bang.
“Why should anyone trust this?”
Nobody on this page asks you to trust its author, which is the point. Every claim carries a link to a source, a dataset, or a test you can run yourself, starting with the standing falsification challenge and its step-by-step protocol. Credentials are a proxy for checkability; this page offers checkability directly. Verify, don't believe is not a slogan here; it is the site's operating instruction.
“Everyone already knew all this.”
Then the record will show it. The related work page maps, line by line and with links, exactly where each component was already established, by whom, and where the joined claim was not. Naming what is old is how this programme proves what is new, and the dated priority record shows when each claim of the theory entered the world. The standing objections, including this one, answered in full.
“It is too hard. The wording is elitist.”
This page was written for a reader meeting every one of these ideas for the first time, and every term it uses is taught on the page in plain words before it is used in earnest. If any sentence here still needs prior knowledge to understand, that is a defect in the page, not in you, and I would genuinely like to know: tell me which sentence.
And for every doubt about the author rather than the work, one answer covers the class: judge the record, not the man. A fantasist does not publish the conditions that would kill his idea; here they are.
A boundary, stated plainly: this page separates its registers and keeps them separate. Religion appears here only as an observed convergence in the human record, never as doctrine, and never as a claim about any particular creator, mythical, divine or emergent. The compounding is observed: self-correcting loops turning intelligence into structure, everywhere they run. The exponent, the pattern and the stability condition are measured and measurable, and they are the theory's only scientific claims. The far horizon is a possibility, recorded as a possibility, and nothing on this estate treats it as more.
Why I am doing this
I did not write this because I was invited to. No institution asked for it, no salary depended on it, and no business plan preceded it. I wrote it because my mind runs every idea it meets through everything else it knows, over and over, until either the idea breaks or the pattern holds. This one held. And when I mapped the possibilities honestly, the cost of being wrong was my reputation, while the cost of staying silent, if I was right, belonged to everyone. That is the whole motive. I judged the importance greater than the risk, and I acted.
I knew what I was claiming was large, so I did what a serious claim demands. I timestamped the core of it to myself in December 2024, before a word was public. I set myself a deadline to publish the book and met it, finishing on New Year’s Eve in the last hours of 2025. I then did something harder than publishing: I began registering, in advance and in public, exactly how my claims could be proven wrong. When one of my own headline numbers failed replication, I retracted it publicly and rebuilt the instrument so the same failure could never happen silently again. The retraction is not a wound in the record. It is the proof the record works.
Before any of it was public, I tested the idea against the harshest critics I could construct, adversarial versions of the people closest to me, built to attack it with everything they had. A version passed only when they fell silent. That practice never stopped: today it is a registered programme of falsification conditions, adversarial review and preregistered tests designed to be able to kill my own claims.
I should say plainly what I am not. I am not credentialled in this field. I am neurodivergent, diagnosed because I went looking for the truth about my own mind, and I have lived my whole life inside the capability fallacy: the assumption that because someone can do remarkable things, their difficulties must not be real, and because their difficulties are real, their work must not be. Both halves are wrong, and both halves are why I build everything on dated evidence rather than on anyone’s benefit of the doubt, my own included.
So I am not asking to be believed. Belief is not the currency here. The claims are published, the tests are registered, the dates are anchored, and the challenge is standing: prove me wrong, and I will publish it myself. Whether I am right will not be decided by anyone’s opinion of me, and it will be decided soon. Check the dates. Run the tests. That is all I ask, and it is everything.