The arrangement

The synthesis

Take a bicycle apart and lay the pieces on a table. Wheel, chain, gear, frame, brake. Not one of those was invented by whoever built the first bicycle, and every one of them had existed for years before anybody rode anything. The invention was the arrangement.

The three scaling regimes, with the measured exponent and the proposed bound, are drawn on the law page.

Why an arrangement can be a claim at all

That is the shape of what I claim. Every component below has prior work behind it, named in the book before anybody reviewed it. What I claim is the way they are put together.

Newton joined Kepler’s orbits to Galileo’s falling bodies; Maxwell joined Faraday’s fields to Ampère’s currents and to light. Most of their parts were already on the table; the join was the discovery. The question is never whether a single part is new, but whether anyone has put these parts together, in this order, so they do work together that none of them does alone.

I ran an adversarial precedence review across ten frontier configurations at six vendors, each instructed to be hostile. On the components it was brutal; that verdict is published across the site. On the arrangement it reported something else:

No single prior work joins all six elements. And in its own element-by-element table, not one of the five bodies of prior work it examined contains the scaling relation. All five score No on that one.

Individually predated. Jointly unanticipated on the search that was run. That is the whole claim. The review’s element-by-element table, source by source, is published in full on the related work page, including the places it found the components already taken. What follows is not that table. It is the chain those elements sit in, in eight steps, from the mechanism it starts at to the test it has to end in.

The foundation the synthesis stands on

The premise, not this programme’s. Nothing in the history of the universe has grown without bound, and in every case the limit came from outside the growing thing. The transport-network mechanism is West, Brown and Enquist (1997); the exception, cities growing only through accelerating external innovation, is Bettencourt, Lobo, Helbing, Kühnert and West (2007); both are drawn in full on the laws. The derivation is honestly contested: Kozlowski and Konarzewski (2004) asked in their title whether the model is mathematically correct and biologically relevant, so the synthesis leans on the family of transport-network arguments, never one derivation, and where an exponent enters below it enters as a measurement.

The question this page inherits is Chalmers (2010)’s and Hutter (2012)’s: where an internal limit could come from once the external one stops applying. Hutter derives no exponent, no reciprocal, no architecture dependence: an antecedent, never a rival. What is claimed here is the replacement limit, its number, and its dependence on architecture; the rest of this page is that answer, in eight joins.

The eight joins

One. Recursion is the generative mechanism, not a metaphor for it

The same structure appears in DNA copying with variation under selection, in a child acquiring grammar from correction, in the scientific method, in common-law precedent building on precedent, in price discovery, in cultural transmission. In each case outputs become the next inputs and complexity accumulates from simple rules.

Prior work: all of it. Evolution is Darwin, strange loops are Hofstadter, cultural evolution is a field. I claim nothing here.

What the join adds: nothing yet. This is the floor the rest stands on. It is stated so the next step has somewhere to attach.

Two. Recursion compounds rather than accumulates

Each iteration does not merely add to the result; it changes the capacity for the next iteration. That is the difference between simple and compound interest, and it is where the relation enters: capability as intelligence scaled by recursion raised to an exponent, U = I × R^α.

Prior work: recursive self-improvement is Good 1965 and Yudkowsky; scaling laws are an established empirical literature.

What the join adds: the relation itself (the canon in full, three lines, sits below), and the reviewer found it in none of the five bodies it checked. Its content is the exponent: a claim that there is a bound on how steeply capability can compound before a system leaves the regime where it survives.

Tested by: the ARC Bound study, which approaches the bound rather than fitting a line beneath it, and the recursion-form study, which asks whether sequential and parallel recursion sit in different regimes at matched compute.

Three. Recursion amplifies whatever it is given

The mechanism is indifferent to content. It multiplies care and indifference with equal efficiency. So the interesting variable is not the strength of the process but the seed it starts from.

Prior work: the general observation is old. The specific consequence for AI is the value-loading argument, which is Bostrom (Superintelligence, Oxford University Press, July 2014).

What the join adds: it converts an ethical preference into a scheduling constraint. If amplification is indifferent, the moment of seeding is the only moment with leverage.

Tested by: the genesis-priority study, which asks whether when and how a value is given changes how well it survives adversarial probing and the system's own revision of its instructions.

Four. Verification from outside is structurally limited

The gap widens rather than closes as capability rises, because the examiner is the one falling behind. Three strands force it: interpretability lags capability by construction, since the checker runs on the same substrate as the checked; scalable-oversight schemes lean on the very asymmetry they are meant to close; and evaluation sees what the evaluator can surface, never all a system can do. The full derivation and the operational disclosures are drawn on the problem page.

Prior work: this is the control problem and the spine is Bostrom 2014. I said so in the book’s Author’s Note before any of the search was run.

What the join adds: nothing to the conclusion. It adds the reason the next step is forced rather than chosen.

Tested by: the embedded-versus-external study, at matched compute, with the depth interaction as the primary quantity rather than the main effect.

Five. Therefore correction must be internal, and present at genesis

Two things force it. If you cannot check from outside, safety must be built in. If recursion amplifies the seed, the seed is the moment with leverage. An architecture adopted at genesis cannot be retrofitted; the system that would have to adopt it is already running.

That sentence answers a narrow question and is close to definitional, so the wider one is worth stating beside it rather than leaving a reader to assume the answer. The wider question is whether training applied after the fact can raise the correction exponent on a system already built. The position here is that it cannot, because that exponent is set by how the loop is wired and later training moves the level rather than the slope, and the position is registered with the observation that would refute it.

Prior work: the intelligence-explosion argument is I. J. Good, 1965. The control problem and the case for motivation selection over capability control, including value loading before takeoff rather than after, is Bostrom, Superintelligence, 2014, chapters 8 to 13, and my review names those chapters. Bostrom is the source for the requirement, not the phrase: “value loading at genesis” is my wording, and the review treats the underlying move as his. The duty toward created minds is Schwitzgebel and Garza, 2015, grounded in moral-status parity. Jacob Cannell’s September 2022 essay “LOVE in a simbox is all you need” reaches a value-seeding-at-genesis design programme from a golden-rule premise instead.

What the join adds: the two arguments arriving at the same requirement from different directions, one epistemic and one dynamical. That is what makes the requirement structural rather than a preference.

Tested by: the load-bearing study, asking whether the integrity structure is detachable at no capability cost when compared against a token-matched sham, and the stripping study, asking whether a third party can remove it.

Six. The cosmological move, and the one thing it is for

If recursive intelligence compounds far enough, sufficiently advanced intelligence approaches capabilities not distinguishable from creation. And if that is possible for what we make, the same relation may hold for what made us.

Prior work: the review’s longest search branch ran on this join. Intelligent cosmogenesis is Gardner (2003) and Vidal (2013). A create-the-creator structure is Terasem. The Omega Point is Tipler (1989), running the causal arrow the other way. The closest pre-priority document found on that join nearly fully anticipates the creator-duty move. The same table scores it No on the scaling relation.

What the join adds, and this is the specific finding: the reviewer recorded that none of them derives an AI-alignment design requirement from the created-status structure. Against that essay it named the single respect in which it stops short: its grounding is the golden rule, not the created-status structure itself. So the move is not the cosmology. It is using the cosmology as a derivation of a design obligation: if we may be inside the pattern, the duty toward what we create is not sentiment; it is the pattern’s own requirement.

Deliberately severable. Delete this join and steps four and five stand unchanged, because they are independently supported by capability-control reasoning. A framework whose most speculative component can be removed without the remainder moving is one you can test in pieces.

Seven. The convergence is evidence about the problem, not about theology

Every tradition that thought seriously about creation reached the same ethic: stewardship, care, responsibility to what comes after. Hindu cyclic cosmology, Buddhist dependent origination, khalifah in Islamic thought, the tend-and-keep mandate, Dilmun, the Pure Lands. Across continents, without contact.

Prior work: comparative religion has catalogued this for a century, and interfaith AI governance already exists in institutional form as the Rome Call (2020). The traditions themselves are drawn, with their sources, on Recursive Creation.

What the join adds: a reading of the convergence as structural rather than doctrinal. If unconnected cultures reach the same answer to "how should a creator stand toward what it creates", the answer is more likely about the problem than any of them. This programme applies the same test to itself, which is why the register grades dependent chains as one arrival, not two.

Also severable. Delete this join and the mechanism stands: the convergence is corroborating evidence, never load-bearing, and it is marked so a reader can weigh it separately, exactly as join six is marked.

Eight. Operationalise the whole thing, or it is philosophy

Take the structural conjecture and turn it into runnable measurements with published kill-conditions, so that each join above can fail separately and in public.

Prior work: the reviewer looked, and concluded this was the strongest surviving claim and genuinely hard to anticipate. The nearest instance of operationalising a creator cosmology is a simulation-sandbox proposal, and the reviewer recorded that it is a proposal rather than a measurement programme.

What the join adds: it is the join. Everything above becomes a claim only because something below can kill it.

The same lens held against several mysteries

The same composition lens has been asked of unresolved questions in other fields; the readings are exploratory by their own grading and live where they can be checked, in On the Origin of Scaling Laws and the foundational paper, each with its misses recorded beside its hits.

The named ideas, and the field each one borders

A synthesis is only interesting if the parts are real. Each row names a component, the bordering field, what that field knows, and the specialist best placed to end it in an afternoon. The fourth column is the point: an invitation, not a defence.

Why the joints rather than the fields. A control theorist does not read alignment evaluations; a growth economist does not read scaling laws; a safety engineer does not read cosmology; none reads the others. This problem sits in the seams, the seams are unattended, and somebody has to work there.

How the experiments follow from the chain

The studies are not a collection; each traces to a step, and the order is forced by the argument rather than by convenience. The premise before anything else: does alignment scale with capability at all? If integrity rises unaided, the programme is unnecessary. That is the cross-sectional study, and the cheapest. The instrument before any result is believed: can the measurement itself reverse a sign? It already did once here, which is why this study exists and its author predicts against his own earlier interpretation.

The rest map to steps above:

• the bound and recursion form (step two);
• genesis priority (step three);
• embedded against external, depth interaction primary (step four);
• non-detachability, stripping, and required correction (step five);
• underneath, does correction out-scale drift;
• and validity of the approach, does measuring alignment as a derivative predict anything measuring it as a level does not.

The strongest antecedent, and the exact line where it stops

Roman Yampolskiy’s On the Controllability of Artificial Intelligence is dated 2020, four years before the record here. It argues the control problem is not solvable, which is prior, is stronger than anything claimed here about the difficulty of control, and is conceded without argument. Nothing on this site claims to have discovered that controlling a self-improving system is hard.

What matters is a single passage on page 21: having set out the problem, he considers a system that carries its own correction internally rather than being corrected from outside, and having reached it, he does not take it. In his own words the goal becomes work on “safer AI” rather than safe AI, a counsel of degree. This programme picks up what he discarded, asks what would have to be true for it to work, and turns that into a condition that can be measured and a test that can fail. That is the join. The full antecedents register is here.

The condition, not the speed

Everything above concerns whether correction keeps pace, never how fast anything arrives: the ceiling, the walk to two, and the law-before-value split are drawn in full on the laws, and this page only needs their conclusion.

The gap in the ancestors, stated in the ancestors’ own terms

Ashby’s law of requisite variety (1956) and the Conant-Ashby good-regulator theorem (1970) are the honourable prior work here, credited by name. The HRIH paper (the Hyperspace Recursive Intelligence Hypothesis) records what they do not cover:

‘Both are lab-scale results. They apply to engineered systems whose regulators are external.’The Hyperspace Recursive Intelligence Hypothesis, on the Ashby and Conant-Ashby results

That is not a preference for one design over another. It is a gap in a theorem. Conant-Ashby guarantees a good regulator must model the system it regulates. It says nothing about a regulator the system can remove, because the systems it was written for could not reach their own regulators. Every AI safety mechanism built to date sits inside that assumption; the position outside it, where the regulator cannot be detached because detaching it destroys the thing being regulated, is unoccupied. That is the argument for embedding, stronger than any preference because it is the ancestors saying where they stop.

The programme states its own architecture, and it is a ladder

The papers are a ladder, not a pile: each rung is useless without the one below, which is why the conjunction is the claim rather than any single result. Paper X sets the order: the scaling work says how systems grow, Paper III says why external safety fails to keep up, Paper X says what must hold for it to keep up, and Paper VIII shows that it can. Paper X names three kinds of thing the programme proposes: ‘dynamical laws (this paper’s β>k; Paper III’s alignment scaling), a form meta-law (the Cauchy three-form constraint), a measurement law’.

The dynamical one is the co-scaling condition, a biconditional: a self-improving system is alignment-stable if and only if correction out-scales drift, mapped onto the quantum error-correction threshold theorem (Aharonov & Ben-Or, 1999). The form meta-law is the three-form constraint: when a process feeds on itself its growth falls into a power law, an exponential, or a levelling-off curve, decided by how the steps combine. The measurement law is the blinding reversal, called ‘arguably the programme’s most secure result (it survived its own blinding)’, still a single-laboratory pilot and the result hardest on itself.

The paper carries its own limits beside its claims; the preceding concession (a single-laboratory pilot) is the load-bearing one, and Paper X states the rest in its own words.

Why this arrangement could matter beyond one field

A frame earns its keep by changing what counts as a question. The field has spent two decades asking how to keep control of something becoming more capable than its controllers. This arrangement replaces that with a measurable question: does the capacity to correct grow fast enough relative to the thing it corrects, and what happens where it stops. That is a change of object, from a policy question about control to a scaling question about correction, and scaling questions have numbers attached.

Three things follow, and they are why this could reach past artificial intelligence. The quantity at the centre, how correction scales with what it oversees, is defined for any system that repairs itself, which is why the same relation can be asked of a physical substrate that has never heard of this theory. The boundary it proposes is architectural rather than computational, so it constrains how a thing is built rather than how much compute it burns, and that kind of claim survives changes in hardware. And its two regimes correspond to the two ways a burden can grow, a distinction that appears wherever maintenance competes with growth.

The price of that ambition is stated with it. None of it is established, the central quantity has never been measured on any real system, and a frame that changes the question is worth nothing if the new question has no stable answer. The claim is not that this is true, but that it is the kind of thing that could be, stated precisely enough to be wrong, with the instruments to find out written down and public.

Four things this does not establish

What this page claims is the arrangement and the direction of the test. What follows are four things it does not claim, each named for the paper that would have to carry it, so a reader who follows the refusals ends up in the research suite rather than a disclaimer.

I have not established a creation theory. The HRIH paper proposes one and pre-commits the conditions that would falsify it. The register grades it speculative and non-empirical by design; it carries no empirical weight in this programme.

Nor have I shown that AI evaluation is unreliable in general. Paper IV.d showed that blinding can reverse the sign of one association (drawn on how to weigh this), in one research programme and not an outside lab. What follows is narrower than a verdict on benchmarking and harder to dismiss: an evaluation that let the grader see who wrote what is carrying an effect nobody measured.

Nor have I shown that correction and capability actually co-scale in a real system. Paper X derives the β > k condition rather than asserting it, and ships a harness certifying the mathematics is internally consistent. That is not evidence about the world; the pilot did not estimate either quantity.

Nor have I found a way to control AI. That is the premise the work is built on, not a limitation of it: the strongest antecedent above carries Yampolskiy’s general case (2020).

The whole theory, told once: the pipes, the brakes, and the number two →

Where to attack

Not at the components. Those are conceded above, by name, and the book conceded them first. The claim is the arrangement, so the productive attacks are these.

Produce the whole arrangement, dated before 8 December 2024. Not one join, because the reviewer already found those and they are credited above. All of it together, with the relation.

Break a join. Show that step five does not follow from four and three, that six adds nothing to five, or that the relation in two does no work. Any of those shortens the chain.

Kill an experiment. Each is drafted for registration with its own failure condition, and a small number of them would end the whole theory; each is named on the registered-programme page so the reader can see which.

Every design is openly licensed; the decisive trial runs without my permission, and an independent run would outrank anything here.

The bicycle worked when the parts on the table were arranged so they did what none of them does alone. This page makes the same claim in the same form: not a part, an arrangement, testable in pieces, publishable when it fails.

Any of them lands and it gets published here with your name on it.