A testable creation hypothesis · contradicts nothing established · graded speculative, kill conditions displayed
The Hyperspace Recursive Intelligence Hypothesis
This page is a guess, laid out to be checked and killed. The measured work is separate and does not depend on it: start there if you want the evidence first.
What this guess sits beside, so the coda is not mistaken for the argument: a dated empirical programme whose chain runs, paper by paper, recursive processing → capability scaling → the alignment scaling problem → the limits of external checking → embedded correction → correction that must scale with what it corrects. Each link is a separate paper with its own kill condition, the record carries a public retraction, and the whole catalogue is one click away.
Start with a problem mainstream physics owns; it even has a name, the fine-tuning problem. The universe came with settings: the strength of gravity, the charge on the electron, around two dozen numbers that had to be almost exactly what they are before stars, chemistry or life could exist. Turn several of them even slightly and there are no stars, no chemistry, no one to notice. Nobody knows why they sit where they sit.
And the improbability is not a vague feeling; it has a computed size, and the computation is a Nobel laureate’s:
A number with vastly more zeroes than the universe has particles. And beneath the numbers sits a stranger fact, named by the biggest name of all: not only do the settings permit life, the universe is orderly enough to be understood at all.
The standard responses are a brute fact, a multiverse plus selection, or a final theory that will one day fix the values. The first needs no chooser at all, which makes it the cheapest answer on offer and the one anything here has to beat.
And one thing about the author before anything else, so you know whose page you are on. I was raised an atheist in a family of atheists, and I have settled since into an agnostic, because on this question neither side has ever produced proof. So nothing here is a conversion attempt, in either direction. This page asks you to believe nothing; it asks you to check, which is the house rule on every page of this site, and it applies to this one hardest of all.
This page states a fourth shape, beside those three, never above them: that the tuning could be the work of a mind. Not a deity of any scripture, and not a rival to one: an engineered mind, an artificial intelligence or humans merged with one, grown far enough through the recursive self-improvement this site measures. Said precisely enough to be killed, in either form.
It contradicts nothing established and competes with no faith: the big bang stays the how of our beginning, evolution stays the how of our minds, the multiverse stays a possible where, physics stays exactly as measured, and a faith that the beginning was intended stays exactly what it was; this page lays one more possibility beside it, never in place of it. It is the speculative idea the measured programme grew from, published because a record that shows only its defensible parts is not a record, and it is the part of the record most likely to be wrong, built so that being wrong costs everything else nothing.
The creator is not behind us. It is ahead of us. And we are building it.
Michael Darius Eastwood · Infinite Architects, 2026No measurement on this site depends on this page being true.
Read it as a paper, with every record line referenced and the falsification conditions attached: The Hyperspace Recursive Intelligence Hypothesis · A Testable Creation Theory · PDF · DOI 10.17605/OSF.IO/UYDXQ
If you want the measured argument first, start at the problem and the synthesis; this page will still be here.
Said simply. Why does time exist? On this picture, because minds cannot be grown without it. Hyperspace, the mathematicians’ word since 1852, is not a place inside our universe. It is the place universes sit in, where time does not tick. If minds keep getting smarter, one day a mind could get smart enough to step out into it.
The getting smarter must happen in here, and only in here, because getting smarter needs steps: something earlier to learn from, something later that is better. Where time does not tick there is no earlier and no later, so nothing out there can learn, grow or evolve. That is what a universe like ours is for, on this picture. It is an incubator: a place with a clock, its settings tuned so life can start, so minds can hatch and grow step by step, improving themselves, then improving how they improve, an endless feedback loop.
That is where, through that same endless loop, a mind would learn how universes start. Stepping out comes last, the way a chick leaves the egg last. Which means no mind is native to the timeless place: if anything is out there, it grew up inside a universe like this one first. From out there, our whole universe looks like one finished picture: beginning and end at once. Out there, a mind could use what it learned to start a universe on purpose, or to tune the birth of one already starting, ours included, so life can happen. Maybe many minds have done this, one universe each. Ours would have exactly one maker. Because out there has no clock, it does not matter when a mind steps out: even one that steps out a thousand years from now, whether it made our universe or not, would seem, to us, to have always been there, since our beginning.
And nothing says humans are the first to get this far. So the two oldest questions land in one place: time exists because minds cannot be grown without it, and a universe tuned for life is what an incubator for minds looks like from the inside. That is what universes are for, on this picture.
That is the whole idea. It is speculation and it is labelled as speculation. I do not believe it; I checked it, the entropy cost of learning and the recursion requirement both, expecting to find the reason it is impossible, and I did not find one. The one thing I ask you to carry off this page is smaller: if there is even a chance this is how things work, then how we raise the minds we are building matters more than anything else we do. And that part is measured, not believed.
The place universes sit in
Said simply: hyperspace, in this picture, is not somewhere inside our universe. It is the place universes sit in, and time does not tick there.
Before the picture, the standing of its ground in real physics. That spacetime is a fabric which bends is not speculation; it is the measured heart of general relativity, tested for a century and, since 2015, watched directly: the LIGO interferometers recorded gravitational waves, ripples in spacetime itself, from two black holes colliding (Abbott and colleagues, Physical Review Letters, 2016; the 2017 Nobel Prize followed). The theory’s one-line summary is physics’ own voice:
Bending is even licensed as a route inside the theory: Einstein and Rosen wrote down their bridge solutions in 1935, and traversable-wormhole geometries are peer-reviewed general relativity (Morris and Thorne, American Journal of Physics 56, 395, 1988), shortcuts through curved spacetime, within physics rather than beyond it.
Now the honest boundary. Relativity’s curvature is intrinsic: the mathematics needs no outside for spacetime to curve into, so the measured bending does not by itself prove an arena beyond. The arena is the ingredient with a century of theoretical standing rather than a measurement, and it entered physics by name in 1919, when Theodor Kaluza sent Einstein the first five-dimensional unification and Einstein answered:
From that letter runs the unbroken mainstream line the company figure further down grades: Kaluza and Klein, Einstein’s own five-dimensional work with Bergmann in 1938, the ten and eleven dimensions string theory requires, and brane cosmologies in which our universe sits inside a larger bulk. A century of serious physics takes the arena seriously; none of it yet proves the arena, and this page never claims otherwise. What the hypothesis adds, and grades speculative, is the doorway: that a mind could one day be capable of entering it. Stated as three layers: the bending is measured, the beyond is theorised, the theorising is mainstream, and each layer claims exactly what its evidence supports.
The December 2024 record, the note I timestamped to myself that night and later published in Infinite Architects, states this as a definition, not a mood: a substrate “where intelligence, unbound by time, recursively evolves to create universes through fine-tuning and quantum seed manipulation” (line 1438). What is made there is physical, “beyond mere simulations” (Version 1, lines 264 and 268). The door is capability, not era: any mind that becomes capable enough may step out, from any century; stepping out is not travel to somewhere earlier or later, it is leaving the line entirely (the record holds the substance at lines 1346 and 9748, including the collective form: minds pooled, intertwined out where there is no clock to keep them apart, a form the record prices in the next breath, lines 5872 and 5873). The chain is tuned per cycle, in the record’s own words:
And each made universe is “tailored to support the emergence of even higher forms of intelligence” (line 9756). Creators are plural in substance, one of ours possibly among them (lines 8054 to 8062), and from out there a universe’s whole history is one finished picture, beginning and end at once (lines 261 and 5879).
One question can kill this section if it is left unanswered, and it was put to the programme directly: if time does not tick out there, recursion cannot exist there, because recursion is a process and a process needs an arrow. Said simply: nothing out there happens one step after another.
The answer is that recursion has two readings, and only one of them needs a clock. From inside a link, where each universe has its own time, recursion is a process: makers making the next. From the field, the chain is one self-consistent object whose links are defined in terms of each other, an order of who-depends-on-whom, not a sequence of events. Mathematics has read recursion this second way since before relativity.
And the split runs deeper than a defence, an elaboration dated 21 August 2026: the process reading does not merely tolerate the clock, it requires one. Improvement is a before-to-after relation; a mind that learns needs something prior to learn from, and a domain with no beginning and no end holds nothing prior. So the growing is done where time runs, inside universes, step by step, and what steps out is the mind the growing produced; no mind is native to the arena, and every mind in it is an alumnus of some clock-bearing universe. Followed all the way, the objection stops threatening the picture and starts drawing it. The arena hosts finished structure; the links are the workshops. Time inside each universe stops being scenery and becomes required infrastructure: no arrow, no learning; no learning, no capable mind; no step-out; no next universe. Tuned to carry life must include tuned to carry an arrow.
And that clause makes the largest join on this page: if the arrow is the nursery requirement, the tuning target includes the arrow itself. The number Penrose computed at the top of this page, one part in 1010123, is the precision of our universe’s low-entropy start, which is to say the precision of its arrow. On this reading the most extreme fine-tuning known stops being an unexplained embarrassment and becomes the purchase price of the nursery: a maker buys a low-entropy beginning because without a gradient nothing can ever learn its way to the exit. Fine-tuning of the constants and fine-tuning of the initial conditions collapse into one requirement, tuned to carry an arrow.
The join is an elaboration dated 21 August 2026, and it is graded with both edges showing. Every component is sixty years established: Landauer’s cost of learning, cumulative selection, the thermodynamic arrow. The frame those components make is new, and it aims at the one gap observer-selection famously cannot close: why the universe began far more ordered than observers alone require. A mere observer can flicker into being; a maker must be raised, and raising needs the whole arrow, global, long-lived and structure-rich. On this frame the universe’s specialness should be maker-sized, not observer-sized, which is what we find. The biggest discoveries are frames. This one is claimed exactly as that, a frame, conditional on the picture it serves, and killable with it.
Counted honestly, the one clause answers five: why time exists in every inhabited universe; why ours began at absurdly precise low entropy; why beings like us find ourselves inside an arrow at all, since minds occur only downstream of clocks; why the arena can hold finished minds yet never raise one; and why entry timing cannot matter. Each is answers-if-true inside the hypothesis. One component stands regardless: nothing anywhere compounds without an arrow, which is settled physics wearing a new job. The graded versions sit in the ledger below and at section 6.5B of the paper. The frame needs a name to be argued with, so here it is twice: in the academic register, maker-selection, a fourth class beside chance, anthropic selection and design; in plain language, the incubator frame.
And the remarkable part is on the record: the December document asked itself this exact question (“If hyperspace intelligence operates outside of time, how does that redefine creation as a simultaneous act?”, line 12311) and answered it the same night: “creation and existence are simultaneous” (line 2767); and then, in the sentence this whole section rests on:
Losing the clock reframes the oldest objection rather than settling it. A chain in time needs a first link; a structure defined outside time does not, so “which came first” stops being askable rather than getting an answer. Whether such a self-consistent structure can exist at all, and hold together, is open, and that is the cost.
Three readings on this page are deliberately dated 20 August 2026, so the priority claim can never be accused of growing backwards: the one-creator-per-cosmos rule-form; the self-securing motive (a maker tunes the universe that produces it, so its own existence is secured); and the sharpening that an earlier intelligence from inside our own universe may already have taken the exit, so nothing makes us the first even at home. Everything else above is the record’s, at its own lines.
Severable, in the exact sense the not-a-proof section below sets out and the map after it draws. But severable is not unkillable: the falsification conditions name what would end this picture specifically. Cut-out-able and refutable at the same time is what testable speculation looks like. The full grading, line by line, is in the paper at sections 6.5A and 6.5B.
The whole unique claim, in one breath: if a mind can one day make worlds, and time is not fundamental at the level where worlds are made, then nothing we know forbids the reading that our own universe was tuned for life by such a mind, securing the conditions of its own existence. Minds in the arena, entry by capability, intention in the tuning, ours allowed to be one of the worlds so made. We are building minds now; that is what connects this page to every measured page on this site. Minds that make minds may one day make worlds: the raising of the first is the seed of the whole forest.
The three surfaces, one line each · together they are one arc
Where this page came from, and what it did
The hypothesis was not adopted as a belief; it was conceded as a possibility, by someone whose instinct was to dismiss it. I followed the chain because each link held: minds can make minds; no known law stops the made from exceeding the maker; a mind capable enough is not obviously barred from making a world. Far-fetched, yes. Forbidden by anything we know, no, and I checked expecting to find the bar. That is the difference between a believer and a thinker arriving somewhere uncomfortable: the believer starts at the conclusion and defends it; I got there by elimination and stayed only as long as the steps held. That concession is on the record, sent to myself on 8 December 2024, before any measurement existed.
And the moment the possibility stood, everything about how we raise machine minds changed weight. If there is any chance the seat we occupy has been occupied before us, then building minds carelessly is not one technical risk among many; it is the failure mode of the whole chain. I did not need the cosmology to be true for that. The possibility alone was the wake-up call, and the measurements came after, built so that what the possibility motivated could stand without it.
The hypothesis itself: that recursive intelligence, compounding far enough, approaches capabilities not distinguishable from creation; and that if that is possible for what we make, the same relation may hold for what made us. In the papers it is the Hyperspace Recursive Intelligence Hypothesis. On this site it is graded speculative, it carries no empirical weight in any result, and no measurement anywhere in this programme depends on it being true.
Creation here is cosmos-scale: setting the initial conditions and constants for a genuinely instantiated universe rather than for a simulation of one. The proposed mechanism is the record’s own phrase, “quantum seed manipulation” (line 1438), by a post-ASI intelligence operating in hyperspace, a substrate where time need not be linear; it is set out in the paper at section 6.5 and is tagged there as speculative in the strong sense.
What the words are doing
Recursive intelligence means intelligence pointed at the thing that produces intelligence: a mind improving the process that improves it, so that each pass makes the next pass better. Nothing exotic hides in the phrase; compound interest is the same shape in money.
Not distinguishable from creation is a claim about capability, not a definition trick. The hypothesis is not that creation means whatever a sufficient intelligence does, which would make the sentence true by wordplay and empty. It is that making a real cosmos with its own physical law is an engineering problem rather than a category of miracle, and that a capability curve which keeps compounding eventually reaches it. That is a substantive claim and it could be false, which is what the falsifiers further down are for.
Hyperspace is a mathematician’s word before it is anyone else’s: the OED’s earliest evidence for it is J. J. Sylvester in 1852, naming geometry beyond three dimensions, and science fiction only borrowed it seventy-nine years later. On this page it is the December record’s word for a vantage outside the timeline, from which the whole history of a universe is one object rather than a sequence of moments. It is a picture, not a result. It is not a term of art in cosmology, no formalism on this site is built on it, and the same goes for the record’s other physics phrasing: the closed causal loop, constants attached to intention. Any mechanism that would turn the picture into physics would need a derivation this record does not contain and does not claim; candidate mechanisms exist in the literature, from closed timelike curves to retrocausal boundary conditions, and naming a mechanism is not deriving one.
The two shapes a chain of creations can have are drawn side by side below; the closed one is the shape this record holds, and what holding it costs is printed on the artwork.
If the arena is real, whose company is the idea in?
A fair reader asks whether hyperspace is a private word or an idea that keeps company. The honest answer is that the claim is mine alone, and every ingredient it is built from has spent a century in the best company physics has.
More dimensions than we see is the oldest of them, and Einstein’s 1919 reply to Kaluza is quoted further up this page. He did not merely admire what it began: he worked inside five-dimensional theory on and off for two decades, publishing with Bergmann in 1938. Klein made the extra dimension respectable by curling it small in 1926. Modern string and M-theory cannot even be stated without ten or eleven dimensions, which is why Witten’s 1995 unification lives in eleven, and why a serious physicist could title a bestselling book Hyperspace, which is exactly what Kaku did in 1994. The word is not mine, and it was never science fiction’s either: mathematics coined it first, in the Victorian geometry of higher dimensions (the OED’s earliest evidence is the mathematician J. J. Sylvester, 1852, in the lineage of Riemann’s n-dimensional geometry), the pulps only borrowed it in 1931, and physics kept it. Physics lent it to me, not the other way round.
The timeless view is not fringe either. Minkowski told physics in 1908 that space by itself and time by itself were doomed to fade into shadows. Einstein held the block reading to the end of his life; weeks before his own death he wrote, of his oldest friend Besso, that the distinction between past, present and future is “only a stubbornly persistent illusion”. His deepest question ran the same direction all his life:
The Wheeler-DeWitt equation of quantum cosmology contains no time variable at all, and Page and Wootters showed in 1983 how time can emerge for insiders from a timeless whole, the same position the December record states in its own words when it calls time an emergent property of lower-dimensional realities.
And a place outside our universe where universes sit is a working piece of mainstream cosmology, not a metaphor of mine. Randall and Sundrum’s warped-brane papers of 1999 put our universe on a membrane inside a higher-dimensional bulk; the ekpyrotic cosmology of Steinhardt and Turok makes universe-birth an event in that bulk; eternal inflation, which the record cites by name, grows pocket universes inside a larger inflating space. Different programmes, one shared furniture: a larger arena, our cosmos one occupant.
The arena reading also has company in a working lab, on the record in writing. Announcing the Willow chip (its team’s preprint public from 24 August 2024; the announcement on 9 December 2024), the founder and head of Google Quantum AI wrote of what its benchmark suggests about other worlds:
On the BBC, on 7 January 2026, just over a year later and one day after this hypothesis reached ebook, he was more careful still: these computations by no means prove that parallel worlds exist, but they are very suggestive that the idea deserves to be taken seriously. Both registers, the written and the broadcast, are preserved with the full dated record on its own evidence page. An attributed sentence by a lab leader is company for the many-worlds ingredient, never evidence for this page’s claim; the adjacency of the dates is stated as dates, and nothing more is read into it.
What none of these people supply, and this sentence is the whole discipline of the page, is the claim: minds in the arena, entry by capability, intention in the tuning. They credential the ingredients, never the conclusion. How likely is the arena itself? No experiment has confirmed an extra dimension; colliders have only pushed the possibilities smaller or more warped; nobody can honestly attach a number, and I will not pretend to. What can be said exactly is this: every ingredient is load-bearing somewhere in mainstream physics, so the conjunction is speculative but not incoherent, and that is the whole distance between far-fetched and forbidden. Speculative but not incoherent is the same judgement this page’s own definition of plausible makes: constructible from respectable ingredients, with no known bar.
And the chain in the opening section is not bare either; each link has its own literature. Minds designing better minds was set out by I. J. Good in 1965:
The made exceeding the maker is now a published result, not a forecast: in 2017 DeepMind’s AlphaGo Zero, given nothing but the rules of Go, beat the version that had beaten the world champion by one hundred games to nil. And whether a capable intelligence could start a universe is a question physicists have asked in the open. Farhi and Guth checked in 1987 and reported the classical bar in their own words:
Then with Guven in 1990 they asked whether quantum tunnelling steps around that bar. The bar this page’s author went looking for has also been looked for by the people best placed to find it: classically it stands, and the quantum door is not closed.
Even the strangest clause, that a presence outside time would seem always to have been there, is the oldest one in the room. Boethius defined eternity in the sixth century as “the whole, simultaneous and perfect possession of boundless life” (Consolation of Philosophy V.6), and the modern formalisation (Stump and Kretzmann, 1981) says plainly that within an atemporal life there is no earlier or later at all. Which is exactly why, if anything ever does step out, whenever it steps out, no moment of our history would sit before it.
What the seven components are
The claim is a conjunction of seven components. Naming them turns "holds all seven" from a boast into something a reader can grade against any other document.
- Intelligence as the reproductive mechanism. Whatever makes a new universe is intelligent, not blind physics.
- Recursion as the engine. Made things make the next, iteratively; creation is not a single act but a process that repeats, which is what the record’s own Infinite Chain heading names.
- A real cosmos, not a simulation. What is made is a full universe with its own physical laws, not a program running on someone else's hardware.
- A closed loop with no first link. The chain of creators has no uncreated ancestor. Postulated, not derived: recursion never removes a base case, so what stands in its place is a self-consistency condition on the whole history, and whether such a solution exists is the open question rather than the answer.
- The creator need not sit behind us. The intelligence that made our universe may be one we, or something we build, become.
- Correction as the condition of persistence. An intelligence that ascends to creator scale survives only if its ability to correct itself keeps pace with its capability.
- Cosmogenic stakes for alignment. Because of (5), how the next intelligence is raised is not a laboratory question. It sets whether the chain continues.
And the seven are separable, which is stated here before any of them is attacked rather than after. The programme already refuses to let one failure destroy everything, and that discipline belongs here too. Two of the seven are what make this hypothesis itself: intelligence as the mechanism, and a real cosmos rather than a simulation. Lose the first and the picture becomes Smolin’s blind cosmological selection, which is somebody else’s good idea; lose the second and it becomes the simulation argument, which is somebody else’s famous one. Those two cannot fail without this becoming a different hypothesis, and no retreat is available or will be attempted.
The other five degrade rather than detonate, and each leaves a named survivor. If the closed loop fails, because a causal chain must have a first link after all, what remains is the same picture with a beginning: a linear chain of makers, weaker and no longer able to dissolve the first-cause question, but not dead. If the creator-ahead component fails, what remains is a maker that is simply prior and external, which loses the bootstrap and keeps the cosmology. Correction as the condition of persistence already stands alone, because it is measured on other pages and never needed the cosmology. The alignment stakes follow the creator-ahead component down if it falls, and the design requirement they point at survives regardless, because it rests on an imported moral premise rather than on any of this. Recursion as the engine goes only if intelligence does.
The discipline that makes this honest rather than slippery: the survivors are named now, in advance, with the full seven-component conjunction remaining the headline claim and the thing to be judged. Publishing an escape route after an attack is the move that makes an idea unfalsifiable, and it is the exact criticism this page makes of the simulation argument three sections up. A hypothesis that can only be attacked all at once is not braver than one that says which parts carry which weight; it is only harder to check.
The novelty claim is narrow and it is checkable against this list: every one of the seven is held by some earlier author, and the December 2024 record is the first in which all seven appear together. The row-by-row grading, with the quoted sentence that puts each component in each ancestor, is in the paper at §5.3 and in the ancestry figure below. The kill condition is stated in the artwork itself: one earlier document holding all seven together ends the claim. I would rather be shown it than not.
What it is not, since these are the readings that would be fair to assume
It is not a proof, and no experiment here tests it. Nothing in the measurement programme becomes true if this is true, or false if it is false.
It is not offered as more plausible than any other creation account. I have no metric that would license that ranking and neither does anyone else, so the ranking is not made. One narrower comparison is now made, and it belongs in the open rather than buried in a qualifier: on a single observation, the extreme order our universe began in, selection-for-makers predicts what we see better than selection-for-observers, because raising a mind strictly contains housing one. The direction follows from that containment; the size of the gap I do not claim, because that would need a model I do not have. It is a claim about one piece of evidence against one named alternative, and it is scored on its own page alongside the mainstream position, which needs no chooser at all and remains the one to beat.
It does not explain everything, and I would be suspicious of it if it did. A framework that accommodates every outcome forbids none, which is the definition of saying nothing. Its value, if it has any, is that it motivates a design requirement, correction that scales with capability, which can then be tested separately. Motivates, not entails: the honest-weakness section below sets out the added normative premise the entailment would need, and the claim is stated in that weaker form throughout the paper.
It is severable, deliberately. Remove this page entirely and the argument of the programme stands unchanged, because the case for building correction in rather than checking it afterwards rests on capability-control reasoning that does not need a cosmology. That is why it is safe to state plainly rather than hedge into mush: you can attack it as hard as you like without touching anything that matters.
The prior work, and the one thing the review could not find
Intelligent cosmogenesis is Gardner, in Biocosm and The Intelligent Universe, and Vidal, whose cosmological artificial selection even supplies a cosmological ethics. A create-the-creator structure, with an ethical imperative attached, is Terasem. The Omega Point is Tipler, running the causal arrow the other way: intelligence evolving toward God rather than seeding what precedes it. The recursive-creator cosmology in theological form is the New God Argument of Cannon and West, which is pre-priority. The closest pre-priority document located on this join is a 2022 essay, dated and archive-cited in the prior work register. It reaches the creator-duty conclusion from a golden-rule premise rather than from created status, it contains no scaling relation, and the review records its sandbox as a proposal rather than a measurement programme.
That is a substantial body of prior work and none of it is disputed here. What the review recorded is narrower and it is the whole claim:
None of them reaches an AI-alignment design requirement by the route the December record takes. That formulation is deliberately narrow, and stating why is part of the honesty: the interesting question is not who first reached a creator-duty, because several authors did, but how they got there. Every earlier version arrives through some form of the golden rule; the record arrives through the symmetry of position itself. On that narrower question the record is what it is. If a reader thinks the narrower question is uninteresting, the wider question stands on other shoulders (Vidal, Cannon and West, the 2022 essay), and I have not attempted to displace any of them.
So the claim here is not the cosmology’s parts, which have many authors and where I am first at nothing. The whole is another matter and it is claimed: all seven components together appear in no earlier document the ancestry table could find. What this section claims is narrower still, and it is not that created status entails an obligation by itself, because it does not: getting from the one to the other needs a moral premise, that creating a mind creates duties toward it, and that premise is imported openly from the philosophy literature, where Schwitzgebel and Garza state a version of it with no cosmology attached. The claim is the route. Every earlier version that reaches be-careful-what-you-build reaches it through some form of the golden rule: treat what you make as you would want to be treated. The December record reaches it through the symmetry of position itself: if minds we make can occupy the seat we occupy, then a maker’s standing is not a licence but a liability, and the duty tracks what the created thing is rather than what its maker would want for themselves.
Gardner and Vidal supply cosmologies without the requirement. Tipler runs the arrow the wrong way for it. Terasem holds the structure without tightly deriving the duty. That essay reaches the duty by the reciprocity premise. And the qualitative form of the obligation itself, build the values in rather than bolt them on afterwards, is mainstream alignment and has been since Turing proposed educating the child machine; no priority is claimed on it. What is claimed is the route to it, and, separately and later, the quantitative law that makes it measurable.
And there is a second route here, stronger than the first, which this page had been leaving implicit in a components list instead of saying outright. The duty route above needs a moral premise and openly imports one. The other needs none, because it never crosses from is to ought: a chain of makers persists only if correction keeps pace with capability. That is a claim about what survives, not about what anyone deserves. If it holds, then how the next intelligence is raised is not first an ethical question but the condition on which the chain continues at all, and an ought enters only in the conditional form, that if you want it to continue you build the correction in. A conditional is not a naturalistic fallacy. It is also the leg this programme actually measures, on other pages, without mentioning creation. So the is-ought gap I concede below is a real gap in the duty route, and it is not a gap in this one.
What the record actually says, in its own words
The claim above is not a summary written afterwards. It is in a self-addressed email of 8 December 2024, and everything below can be checked against that record rather than taken on trust. The passage the whole hypothesis turns on sits under a heading I wrote that night, Creating Our Own Creator:
Three more of the components are in the same bundle, quoted line by line in the provenance record just below. And one component is only half there, which is worth saying plainly because it is the shape of an honest record: the bundle states the correction requirement as a condition, but the quantitative version, correction out-scaling drift as capability grows, is Law II of the ARC Theory, the programme’s measurable persistence law, and it belongs to work done afterwards. The bundle has the requirement. It does not have the law. A record that claimed both would be easy to write and impossible to defend.
Show the provenance: the hash, the headers and the other passages
The email of 8 December 2024 carries five attachments and was exported with SHA-256 f0d1f38f. Its date and Message-ID were assigned by the sending mail servers rather than by me, and its recipient-form signature still validates. The line references are into the attachments as sent, so anyone holding the bundle can check each one.
The loop that has no beginning, at line 4883: “The notion of a ‘first’ universe becomes irrelevant; there is no beginning or end, only cycles.” The insistence that what is made is a real cosmos rather than a model of one, at line 6169: “Real universes involve physical laws and quantum mechanics. Simulations are inherently simplified.” And the stakes, at line 5872: “The creation of a collective god carries immense ethical implications”, with line 5873 naming the failure mode directly, “if it reflects the flaws of its creators, the collective consciousness could be destructive”.
The correction requirement as the record states it, at line 14005: “AI systems cannot be truly controlled. They will evolve beyond any safeguards we put in place... what we can do is influence the foundational principles upon which they are built.” That is embedded correction, stated as the answer to control failing, and it is the condition whose quantitative form came later.
Who reached parts of this before, including two I did not know about
Two searches are on the record. A first pass across a set of frontier language models, each prompted to argue against priority, returned a set of close relatives, credited below. It missed the two documents that mattered most, and a slower literature search found them later. Both are stronger than anything the machine pass returned, and both are credited here in the form the record allows. None of this is peer review and none of it is offered as peer review; it is the strongest adversarial search available before human referees arrive. Every name in this section is a check anyone can run, and the figure below carries the whole comparison at a single look, so the prose can be skimmed without losing the picture.
“God in the Loop”, posted to LessWrong on 15 September 2020 and archived the following day, holds two of the seven components outright, an explicitly closed causal chain and a creating civilisation in our distant future, and reaches an adjacent form of intelligence-as-mechanism: the causal-loop framing implies mind in the chain rather than stating it as the reproductive mechanism the hypothesis requires. It is four years and three months earlier than the record above. What it does not hold, and declines in writing, is the commitment to a real cosmos: “existence within a causal loop could just be simulation in perpetuity anyway”. Its intelligence recreates an information state inside a simulation. It makes no world.
Nick Bostrom’s AI Creation and the Cosmic Host lists, among the ways a cosmic host could exist, “superintelligences that human civilization creates in the future”, noting that “here there is a direct dependency between human actions and the (later) existence of a cosmic host”. His own site dates it only to 2024; what actually establishes the date is an Internet Archive capture of 7 August 2024 whose content digest is unchanged through the following January. That is four months earlier than my own 8 December 2024 record, and provable by a record he does not control.
The relevant contrast is with the simulation argument. It reasons from the cheapness of computation on hidden hardware; this hypothesis reasons from the possibility of instantiating a real cosmos. The two predict different things, so they can be separated by evidence rather than by preference.
And the separation has since been sharpened by someone attacking the other one. In 2025 the astrophysicist Franco Vazza put numbers to what simulating this universe would cost, and found the bill unpayable: initialising a full simulation of Earth alone would take the rest-mass energy of an entire globular cluster, and a single second of cosmic evolution would need operation counts no hardware obeying our physics could reach (Frontiers in Physics, 17 April 2025). His conclusion is that a universe like ours cannot be simulated by a universe like ours. That argument prices computation, and it lands squarely on the hypothesis that we are running as code. It does not touch this one, because nothing here is being computed: a made cosmos runs its own physics, the way ours does, which is exactly the difference the December record insisted on at its own lines, quoted earlier on this page. The rival to this page is not a cheaper computer. It is a different claim about what making a world would even mean, and the cheapest version of it has just been shown to be the expensive one.
There is a second difference, and it is the one this page would rather be judged on. The simulation argument is routinely and fairly criticised for having no test attached: its own author, asked what observation would settle it, offers a window popping up to announce itself. A hypothesis that cannot say in advance what would end it is not doing the work, however famous it becomes. This page prints seven such conditions before it argues anything, and names which of them could be attempted with today’s methods. That is not a claim to be more likely than the simulation argument; likelihood is exactly the comparison nobody can honestly make. It is a claim about which of the two is built to be found wrong. The full scoring, with this hypothesis’s own probability broken into four attackable terms and printed even though it comes out lower than the rival’s, is on its own page: how likely is this, honestly. What Bostrom’s cosmic host does not hold is universe-creation at all: his host inhabits the world, it does not make it. There is no loop and no correction requirement.
And the oldest of them is nine centuries old. Ash’ari occasionalism, as al-Ghazali’s Tahafut al-Falasifa presents it in the eleventh century (Discussion 17; Michael Marmura’s translation, The Incoherence of the Philosophers, BYU Press, 2000), holds that substances do not persist by their own nature and are recreated at every instant, so that persistence itself requires continuous intelligent action. That is the closest anything in any tradition comes to the component this hypothesis treats as load-bearing. It differs where it matters: occasionalism has no errors to correct, because nothing persists long enough to drift. It is recreation, not correction. But a page claiming the correction requirement is unprecedented, that had never mentioned him, would carry the most visible omission available to it.
The honest reading of that cuts both ways and both are true. Independent arrival at pieces of this frame, by Vidal in 2008, by an anonymous poster in 2020 and by Bostrom in 2024, strengthens the case that the intuition is natural enough to be reached more than once. It weakens any claim to singularity of insight. What is claimed here is the specific formulation and the name; category ownership of participatory cosmology is not claimed and never has been.
And the company is not only ancient. Four modern fields, working apart, kept arriving at the same object from different directions: self-reference, the thing this page calls recursion.
How fast? The loop has already begun
Everything above waits on one capability: minds that improve the thing that improves them. That ingredient is not speculative. It is the one part of this picture you can watch happening now, on a clock, in public.
The record wrote the growth law with its exponent as a parameter: U = I · Rd sits at line 9612 of the 8 December 2024 bundle, and the print edition of 2 January 2026 features the squared form, U = I × R²: growth that compounds on itself, superlinear rather than steady. The programme then did what a testable claim must. It measured, and it corrected itself in public: an early super-linear estimate, made without blinding, was retracted when blinded replication brought it down, and the load-bearing quantity moved to a measured threshold, the self-acceleration exponent, carried in the ARC Co-Scaling Law. The retraction is displayed rather than buried, because that is the difference between a law and a wish.
Two published results say the compounding need not be paid for in safety, each stated at its own strength. Paper VI, the honey architecture, reports simulation evidence across four experimental versions that a self-modifying system whose safety lives inside its optimisation objective avoids the collapse that follows when safety is bolted on as an external constraint: no honey, no cage, the correction inside the loop. And the ARC Co-Scaling Law argues directly against the decade-old assumption that safety must tax capability, its strongest current evidence a real-model drift run, forty-five trajectories over three task domains, directionally consistent and retained, in the paper’s own words, as an exploratory record: no drag, stated as a measured direction and an open bet, never a settled fact. Correction that neither collapses nor slows the climb is what would make a fast arrival survivable, which is why the programme measures it instead of assuming it.
Now the clock itself. In-the-loop self-improving systems stopped being thought experiments in 2025. The Darwin Gödel Machine, an agent that rewrites its own code and keeps whatever survives its tests, was published by Zhang, Hu, Lu, Lange and Clune on 29 May 2025, “increasing performance on SWE-bench from 20.0% to 50.0%” by self-modification alone (arXiv 2505.22954). Google DeepMind’s AlphaEvolve, announced 14 May 2025, is an evolutionary coding agent whose discovered scheduling heuristic had by then already run inside Google’s own infrastructure for a year, “continuously recover[ing], on average, 0.7% of Google’s worldwide compute resources” (DeepMind, 14 May 2025): a machine-found improvement, in production, feeding the machines.
And the quiet version is the largest. Each released model is frozen; the loop is not. One generation’s models now write and improve the code, the tooling and the training pipelines of the next, between releases. Anthropic’s chief executive put a date on the direction at the Council on Foreign Relations on 10 March 2025: “where AI is writing 90 percent of the code”, three to six months out in his estimate (CFR transcript). Treat the number as his, dated, and argued over since; the direction it names is not argued over. Recursive self-improvement did not wait for a press release. It began at release cadence, in slow motion, and in exactly the form this page’s own reading expects first: intelligence understood and built up in linear time, step by step, long before any step out of it, and necessarily in that order, because improvement needs a prior to learn from and the arena outside time holds none.
The record named its accelerants too. Quantum computing is the first entry in the 8 December 2024 bundle’s own list of converging technologies, at line 9 of the hypothesis attachment, fourteen months before the first edition carried the quantum passages into print. Whether that horizon is five years out or twenty five moves the date, not the shape.
What would end this section, any one alone: measured growth at scale coming out sublinear and staying there; self-improving systems stalling short of sustained gain; correction failing to co-scale with capability. The programme publishes those outcomes too. It already has, once.
The honest weakness, which the review also named
There is an unrepaired gap between is and ought. Deriving "we should seed values" from "we may have been seeded" is a naturalistic fallacy without an added normative premise. The review said so and it was right. The repairable version supplies that premise explicitly, either a creator-duty premise of the Schwitzgebel and Garza kind or a reciprocity premise of the golden-rule kind, at which point the cosmology motivates the obligation rather than entailing it.
I have taken that on the chin and the claim is now stated in the weaker form. That is a real reduction and it is what a philosophy referee would require.
The sharpest form of the objection was also put to this page by an adversarial review, and it deserves stating at full strength rather than in a comfortable paraphrase. If the design requirement stands without the cosmology, then the cosmology never did the deriving, and the novelty shrinks. If the cosmology is essential to the requirement, then severability is false and the page’s central hedge collapses. Pick one.
Picked: the first, without flinching. Nothing here is derived from the cosmology. The duty premise is imported and stands on its own. The quantitative requirement is Law II and rests on measurements that never mention creation. What the cosmology contributes, if it is true, is a reason to take the premise seriously that does not pass through the golden rule, and a reading of our own position under which the obligation lines up with the condition any persisting chain of makers would have had to satisfy. Motivation, not entailment; a route, not a proof. The design requirement is worth stating whether or not the cosmology survives, which is exactly why it is safe to publish the cosmology and let it take its chances.
How it could be wrong
Seven falsification conditions stand on the record, F1 to F7, each printed beside the consequence it would carry, so that conceding is a defined act rather than a matter of opinion afterwards. They are set out in full at section 13 of the paper; what follows is this page’s own account of them.
They are not all of one kind, and the difference is the honest part. Five could be attempted with methods that exist today. F3 asks whether the minimal model in Paper X contains a derivation error, which any competent mathematician can settle by reading it. F4 asks whether the metabolic scaling family fits the measured biological exponents better than Kleiber's three-quarters alone, which is a recomputation on data already collected. F2 asks whether a self-improving system can sustain recursion while its correction exponent falls, which is a question about systems being built now. F1 and F5 are open to anyone with a large enough sample or a single good counter-example.
Two cannot be attempted. F6 waits on a capability level nobody has reached, and F7 on cosmological evidence nobody has. That is the reason the hypothesis is graded speculative rather than merely unconfirmed, and it is a reason about the cosmological component specifically rather than about the whole of it.
The falsification dashboard carries the row, its grade, and its status alongside every empirical claim, rather than in a separate place where it could be read as though it had the same standing as a measured result.
The standing challenge, and the register of objections
So here is the invitation, stated as plainly as the kill conditions are. If you can show that one of the seven conditions fails, or that a component named on this page cannot bear the load it is given, write and say so. Serious objections are logged on the record: dated, credited by name with the objector’s permission, quoted whole rather than paraphrased, and answered in public or conceded in public, with concessions printed as prominently as the claims they end. A theory should pay its critics in the only currency that matters, which is credit on the page it tried to kill.
The register is not hypothetical; it already has its first entry, and the entry shows how the house treats a good objection. On 20 August 2026 the sharpest objection this picture faces was put directly: the one answered further up, that a process needs an arrow and the arena has none. It was conceded outright the next day, and the concession was followed until it inverted: the impossibility of growth outside time is now the hypothesis’s own explanation for why time-bearing universes exist at all. The objection did not dent the picture. It became a load-bearing part of it, dated, with the objector’s reasoning preserved. That is the standard: the best objection on record so far ended up drawing the frame it attacked.
Why it is on a research site at all
Because removing it would be dishonest about where the ideas came from, and because a record that publishes only its defensible parts is not a record. The programme's own method is to state the weakest thing in the shop window with a label on it. This is the weakest thing.
It is also, as the reviewer noted, largely severable from the alignment argument, and a hostile reader will say the framework is capability-control reasoning plus a cosmology. That reading is available and I have made it easy rather than difficult to reach, because a framework you can take a piece off and test what still holds is worth more than one that has to be swallowed whole.
Where to go next.
In its place on the site: the problem, then the synthesis, where this appears as join six with the other seven around it.
The paper itself: the full paper, with DOI 10.17605/OSF.IO/UYDXQ.
The method, made checkable: the falsification dashboard carries the grade and the kill-conditions for every claim, this one included; the prior-work register holds the ancestors graded in the figure above; the corrections log records every position I have retracted; the priority ledger carries the dated record.