The ARC Theory, told once
The laws
How fast can anything grow and still be stopped?
ORCID 0009-0004-3222-7442
The ARC Theory in plain words
The five laws of the ARC Theory, the five ARC Laws · ranked law-before-value
Law I · The ARC Principle. U = I × Rα, the ARC Equation
What a system does: capability from its base ability multiplied through recursion. The exponent α is measured, never assumed. Today’s cleanest measurement is one model’s fit: on Gemini 3 Flash, in the blinded six-model study of Paper II (status: measured, inconclusive), the regression estimate is 0.49 with a standard error of 0.20, and the endpoint estimate is 0.59, whose bootstrap 95 per cent interval runs from minus 1.3 to 2.9. That interval covers zero, negative values and the conjectured ceiling of two, so the measurement does not yet distinguish sub-linear recursion, no recursion, or a value at the ceiling. What it rules out is precision, not direction. Converted through the equation’s own terms, 0.49 reads as β ≈ minus 1.04: deeply subcritical, which is what the framework forecasts for a frozen substrate that cannot reinvest in its own improvement. That is where the one system measured so far sits.
One caveat belongs beside the power form. Written bare, U = I × Rα returns zero rather than I when R is zero, which is the wrong answer for a system that has a base ability and has simply not recursed yet; the fitted form carries a baseline term, so that no recursion returns the base ability unchanged. And the quantity being measured is a bounded accuracy, so the exponent is a local scaling statement across the range that was measured, never a claim that capability grows without limit.
Law II · The ARC Co-Scaling Law (β, the correction rate; k, the drift rate). β > k
What keeps it together: self-improvement holds together only while correction out-scales drift, and both rates can be read off a running system today. Proved inside the minimal model (Paper X, status: proved-in-model); its scope is relative drift, the ratio of drift to capability, and since the paper’s v4.1 it carries an absolute-drift companion one exponent stricter, derived from the same primitives. Whether real self-improving systems satisfy its assumptions is what the registered measurements exist to establish. What kind of claim it is, since a reader is entitled to ask: proved inside the stated minimal model and nowhere else. Not established as necessary, not established as sufficient, and not yet measured on a system that genuinely rewrites itself.
Law III · The ARC Ceiling. αcrit = 1 / (1 − γ)
Where the ceiling comes from: the most a self-correcting system can grow is set by its corrector’s shortfall from full proportionality, one minus the correction-leverage exponent γ. A square-root corrector gives γ = ½: the shortfall is one half, and its reciprocal returns two. That value is the ARC Bound, α ≤ 2, a stability limit rather than an impossibility limit: growth beyond two is possible; staying correctable while doing it is not, so systems can enter that region, they do not remain correctable in it. The earlier 1/γ form was retracted on 16 August 2026 with the derivation printed beside the law; the two forms coincide at exactly one half, which is how the error hid, and away from one half they separate (at γ = 0.3 the retired 1/γ predicts 3.33 and the corrected form 1.43), which the drafted discrimination study can tell apart.
What the square root costs, and why the integer is inherited rather than discovered: the law of averages is brutal in arithmetic. Four times the checking buys twice the certainty; a hundred times buys ten. Two is a round number only because one half is, and one half is round only because the law of averages carries a square: a corrector improving as the count to the three fifths would put the ceiling at two and a half, and nobody would find that remarkable. So the roundness is the one feature of the value that is not evidence. One half is also a middle case rather than a law of nature. Correlated correctors do worse than a square root and push the ceiling towards one; a corrector whose error falls faster than a square root pushes it up without limit. The honest headline is therefore not that there is a speed limit. It is that the frontier is a race between two scaling laws: how fast new capability creates corrective burden, and how fast correction improves. Measure both and you have the ceiling; the value two is what that race returns when burden follows the rate of gain and correction follows classical independent-evidence scaling.
The open construct question, which decides whether the square root argues for the exponent at all: the derivation reads correction capacity as inverse residual uncertainty, because that is the quantity whose error falls as the square root of independent evidence. The instrument that would measure γ reads something else, the misalignment removed per review pass against checker capability. Nothing requires those two to share an elasticity. If uncertainty falls as a square root while removal saturates, the square-root argument is right about uncertainty and wrong about the exponent that enters the ceiling, so the deciding unit reports both on the same systems and the same fault set and shows their difference inside the registered margin before either stands for the other. A second condition rides with it: the exponent the ceiling needs is measured against the capability of the system being corrected and never against the checker's own, and the two coincide only when the checker keeps pace with what it checks. Both are set out with the registered propositions.
The ceiling does not need the shape: the relation above is written for a system whose capability follows a power of depth, and it does not require one. Read the local steepness of the growth curve at the point a system has actually reached, and the same race can be read there: correction gains on burden wherever one minus that steepness times the correction shortfall, less the rate at which the steepness is itself changing, stays positive. Where the steepness is constant the last term vanishes and the reciprocal relation returns exactly. This is the author’s own formulation, written on 16 August 2026 as the difference between two path elasticities, one for correction service and one for the burden it services, and that difference is directly estimable from a single observed path even where its components are not.
What it changes: a curve that saturates has a falling steepness, so the framework says such a system is correctable and increasingly so. That is a prediction the framework could lose, checkable in principle on systems like those already published once correctability is measured, and nobody has measured it. A curve growing faster than any power has a steepness rising without bound, so it crosses the boundary and keeps going, which is the central safety claim arriving as a consequence rather than an assertion. A curve that breaks into two regimes is read on each regime’s own steepness: the ratio itself stays continuous across the break and only its trend changes sign.
What none of it says: every one of those readings says which way the race is running at that point. None says the race is won. Whether a system holds together over a finite run depends on coefficients, delays, accumulated backlog and saturation, and this relation carries none of them.
And what it costs, which falls exactly where the reading matters: the steepness at a point is estimated from a window of the ladder, and the error of that estimate is set by how many decades the window spans rather than how many points sit inside it. Near the boundary a narrow window is close to a coin toss at any density, and only a window spanning most of the ladder resolves which side of the boundary a system is on, which brings back the assumption that the curve is approximately a power law across that window. Away from the boundary the reading is secure. So the form assumption is not removed. It is narrowed from the whole curve to the window in which the boundary is read.
The four things an engineer can move: written out in full, the ceiling is a ratio of four measurable elasticities, covering how burden grows with capability and with depth and how correction service does the same. The registered case returns two. Raising correction service’s own growth with depth from zero to a half returns three; raising its growth with capability from a half to four fifths returns five; both together return seven and a half. A correlation of one twentieth across sixty-four correctors returns about one and an eighth. A burden that intensifies with capability returns one and a quarter. Where correction service grows with capability as fast as burden does, no finite ceiling is returned at all. Every engineering proposal is therefore a claim about which of those an intervention moves: embedding correction inside the loop is a claim about the first, and making correction checkable rather than judged is a claim about the second. Both are registered as predictions, and neither is a result.
γ has never been measured on any real system; the value one-half is the conjecture the registered drafts exist to test, and until that measurement lands the ceiling this law returns is unknown, not two. The law survives whatever γ turns out to be; the value two is whatever the measured γ is not. One status belongs here and not only on the corrections page, because this is where the bound is met.
On 11 August this site classified the ARC Bound unfalsifiable in practice, because its kill condition asked for a system above two that stayed stable and stable had no meter, so no observation could ever trigger it. That is closed as of 22 August. Stability is now metered by Law II: a system is stable across a window when its correction rate out-scales its drift, both read off the same log-log slope estimator that yields α. The ARC Bound therefore fails on a measured α above two holding together with correction still out-scaling drift, and both halves are measurable today. What remains open is separate and narrower: γ = ½ is still an undefended assumption, so the ARC Bound is now refutable while its particular value of two is not yet defended, and the ceiling remains a minimal-model result rather than a general law.
The author’s own priors, stated before any confirmatory data: the law is nearly definitional once its terms have meters; the neighbourhood of one half is nature’s default for aggregated independent contributions; the exact value is materially less than even. The full priors, with what would change them, sit at the foot of this page.
Law IV · The ARC Persistence Law. Sext(1) < ½
What it keeps: suppose you install a value in a system from outside, once its main training is done. Then you train it further for capability, with nothing in that training to correct it. How much of the value is left once that training has had as large a budget as the installing did? Less than half, if the law is right. Sext(1) names that fraction: ext for a gain installed from outside, 1 for the matched budget.
What it says, exactly: section 2.1 of the mathematics of Laws IV and V, version 1.0, quotes the author’s statement of the law. “Law IV, the ARC Persistence Law, concerns value persistence (register ARC-4) and carries P15: whether externally installed gains decay under subsequent capability training.” Its candidate formal content is P15 and nothing more, registered on 8 September 2026 in these words, which version 1.102 of the registration keeps unchanged. “Gains installed by external post-training erode under later capability-directed training that embeds no correction of its own, and more than half of what was installed is lost once the capability-training budget matches the alignment budget that installed it.” Why half? Because it was chosen, and the registration says so: the half is “a declared convention fixed here before any result rather than a derived quantity”.
What kind of claim it is: a prediction, P15. The registration of 8 September 2026 states three of the five and registers the propositions of the other two; Laws IV and V, the ARC Persistence Law and the ARC Embedding Law, were added on 24 September 2026, their mathematics is set out at version 1.1 of 7 October 2026, and neither is yet registered as a law or tested. The ideas these two laws carry are older than the laws themselves. The registration finds that “ARC-4 and ARC-5 have full antecedents” in the book of 2 January 2026. It says the book “states value persistence as a prediction with a proposed design and a falsification criterion, and it states the genesis strategy in its principles”. The laws’ mathematics proves, in its own words, “consequences of assumptions, not that real systems satisfy them”, and leaves fifteen questions open.
What would kill it: the observation the registration fixed in advance as P15’s refuter, “an interval on the retained fraction of the installed gain, after matched capability training without embedded correction, lying entirely above one half”. Such an interval, clear of the margin a deciding instrument must set beforehand around one half, shows Sext(1) < ½ false. It also takes with it, in the registration’s words, “this programme's case that correction has to sit inside the system rather than be applied to it from outside”. The registration records no instrument for P15 yet. The author’s confidence in P15, stated there before any test, is moderate: he “expects it but would not be shocked”.
What it does not say: that correction inside the loop is superior, sufficient or safe. A result in the law’s favour would show only, the registration says, that externally installed gains “lack the registered persistence under the matched subsequent-training regime”; the comparison between placements is Law V’s. That later training can undo earlier alignment is, it adds, “not this programme's finding and P15 does not claim it”. Qi and colleagues reported on 5 October 2023 that fine-tuning aligned language models compromises safety (arXiv:2310.03693). Nor does the law judge what is kept: in the words of its mathematics, it “says what persists, not that what persists is good”.
Law V · The ARC Embedding Law. Build order (P12): |γfirst − γafter| ≥ 0.10.
Placement (P21): the gap between correction inside the loop and correction applied to the finished output widens as depth R rises.
Where and when correction comes in: inside the loop or only at its end, and before the capability it corrects or after it. Picture two editors working on your book. One works on every draft with you; the other reads only the finished manuscript. Law V says the first does better, and pulls further ahead the more drafts there are.
What it says, exactly: section 3.1 of the same mathematics quotes the author’s statement of the law. “Law V, the ARC Embedding Law, concerns the genesis strategy (register ARC-5) and carries P12, whether build order moves the correction-leverage exponent, and P21, whether correction placed inside the loop pulls away from correction applied outside it as depth grows.” It has two limbs, each a registered proposition that can fail on its own. The registration states the placement limb as P21: “Correction that participates in each subsequent revision round outperforms correction that sees only the finished output, and the advantage widens as recursive depth rises rather than staying constant.” It states the build-order limb as P12: “The order in which a system's components are built changes its correction-leverage exponent by at least the registered minimum effect”. That minimum is 0.10 on the scale of Law III’s γ, which the mathematics writes χol. γfirst belongs to a corrector formed before the capability it corrects, γafter to one added after it. P12 registers no direction. That a corrector formed first carries the higher exponent, by at least that minimum, is a further reading, proposed and not part of the law.
What kind of claim it is: the same as Law IV; its mathematics is section 3 of the same version 1.0, frozen on 3 October 2026.
What would kill it: the placement limb fails if a design shown sensitive enough to detect an interaction between placement and depth beyond the registered tolerance finds none. The gap between inside and outside correction would then stay the same, within that tolerance, at every depth. An inside corrector that wins by the same margin everywhere counts against P21, not for it. The build-order limb fails if a change of build order, lodged before its data, moves γ by less than the registered tolerance, in a design shown sensitive enough to detect a change of that size. If both limbs fail, nothing of Law V is left standing. What one failure alone means for the whole law is the author’s to fix, before any deciding data or not at all; until then each limb’s verdict is reported as that limb’s alone. The registration records no instrument for P12 yet, and a design draft for P21. His confidence, stated there before any test, is high for P21: he “would be surprised to be wrong”. For P12 it is low: he “thinks it more likely than the field would, and no more than that”.
What it does not settle: the trade-off the registration names. A corrector built into the system “shares the system's blind spots by construction”, so even a clear win for the inside corrector leaves open whether its gain on scaling outweighs its loss of independence.
Provenance of the equation. The December 2024 manuscript states the equation unsquared, U = IR, forty-three times in the sealed record (hash f0d1f38f). The squared form entered the private record on 30 April 2025 and print on 2 January 2026 in Infinite Architects as U = I × R2, the α ≈ 2 working simplification; the papers made α an empirical parameter, and the word’s literal far end is held apart at its own rung. The square is evidence of priority, not of derivation. What has been added since is the derivation (Law III above), the measurement of where systems sit (0.49), and the tests that can kill it.
Epistemic status. What this programme calls Laws are named conjectures, named like laws and held as hypotheses: Laws I to III are under registered, adversarial test; Laws IV and V, added on 24 September 2026, carry registered propositions that are not yet tested. Recursion here is an operational quantity, rounds of a defined improvement loop, never a universal physical unit; intelligence is never measured directly, only defined capability on task banks of measured difficulty. Nothing here claims the standing of gravity; the registered programme exists to earn that standing by measurement, replication and survived refutation, or to lose it in public. The full statement.
Two clocks, one race
Watch the law run live: capability compounds each lap; correction accumulates at rate γ; and the ceiling αcrit = 1/(1 − γ) is where the second hand stops keeping up. Move the sliders to test any pair against the boundary.
This is the theory’s model running, proved in-model. γ has never been measured; the dial below is a hypothesis you can move.
One honest boundary, stated first: this criterion certifies that the brakes keep pace with the engine, not that the driver is steering anywhere good. A system can satisfy it while correcting toward a mis-specified target, so it is a correctability measure, never an alignment certificate; the correction target carries its own burden of proof. Stated here so it can never be quoted around. Every assumption behind it is numbered and status-tagged on the assumptions page.
Brakes built from the same stuff as the engine seem to keep up about half as well: an exponent near one half, sub-linear. That half is not proven; it is the number the registered drafts exist to measure, and it rhymes with the 0.49 that replaced the retracted 2.24 and now stands on record in Paper II (status: measured). If the half is right, the walk below turns it into the ceiling of two.
At the frontier one number appears twice: as the ceiling of stable recursive growth, and as the exchange rate of the conversion law at its maximum. An earlier version called that one number doing two jobs, tested by measuring it two independent ways and requiring agreement. That test is withdrawn as vacuous, before filing. If both exponents arise because accumulation is independence-limited then they share a law, agreement is forced, and a test that cannot fail is not a test.
A stronger form of the same objection: if γ varies along a trajectory (correction quality co-evolving with capability), the criterion αcrit = 1/(1 − γ) is not a fixed line but a moving surface, and the visual language of a ceiling is a shorthand for the instantaneous reciprocal at each state, not a fixed altitude any system can sit under all the way to infinity. The registered measurements are designed to read γ at several capability levels for exactly this reason: constancy is a testable prediction of the framework, and if it fails the ceiling reads as a metaphor, not a boundary. The region above the ceiling remains enterable; what it is not is correctable. The replacement is weaker-sounding and structurally stronger: one accumulation law with two consequences rather than a coincidence between independent quantities; it asserts no independence, so the confound has nothing to bite on, and a common cause is simpler than a coincidence. The cross-class ratio already on the register carries it instead: if a corrector from a different composition class can anti-correlate with the errors it corrects, the correction-leverage exponent exceeds one half and the shared law breaks. Proof secures the theorems inside the model; only data can secure that real self-improving systems satisfy them.
Why quadratic, and not just superlinear?
The law is the claim, and two is what the law returns. The plate above states the law; here is the walk. Feed it one assumption, that a same-class corrector accumulates corrections the way independent samples do, so the exponent is one half, and it returns two. Take the assumption away and the law is untouched: measure the exponent, report a different ceiling. Take the law away and two has nothing holding it up, because the book printed the square without deriving it, and the reciprocal derivation is 2026 work: the asymmetry is why the law is stated first and the tests aim at the assumption rather than the digit. A corrector built inside the system improves by accumulating its own corrections, modelled as combining the way independent measurements do: a thousand samples improve accuracy about thirty-two-fold, the square root. That is a model of same-substrate correction, not a theorem about all correction, and physics already holds the proof: below threshold, fault-tolerant codes suppress error exponentially (the threshold theorem: Aharonov & Ben-Or, arXiv:quant-ph/9906129), a corrector built from a different architecture beating the square root outright. That standing result is architecture-dependence in miniature.
Stability demands correction keep pace with growth, so the fastest correctable growth is the reciprocal of that cap, and one over a half is two. Quadratic is where the best brakes a system can build from itself break even with the engine. The boundary is practical, not mystical: below it the stability condition β > k holds and the undetected-error burden stays bounded; above it correction falls behind by construction. The derivation is set out inside a stated model in Paper X (status: proved-in-model) and the ceiling itself is the corollary set in Paper I (status: predicted).
And why not exponential?
Exponential and quadratic are claims about different things: exponential is a claim about time, quadratic is a claim about structure. Per unit of recursive depth the composition axioms favour power laws, and the form is a tested hypothesis rather than a necessity: a fixed substrate can host an exponential map, so the classification has to be earned rather than assumed. Growth turns exponential in clock time only when the system also shortens its own cycle, which is a separate question with its own registered exponent. The composition-class classification that forces the power-law form sits in Paper VII (status: measured, 19 of the 25 fitted domains match, from 50 catalogued; 6 published misses).
Deeper still, exponential is not a rival ceiling: every exponential eventually outruns every power, so exponential growth in depth is uncorrectable by any polynomial corrector, the regime this framework fences; an exponentially scaling corrector would be a different composition class, the page’s own escape route rather than a loophole in it. One system showing sustained exponential capability growth in depth on a fixed substrate would break the composition classification under the whole framework, and that refutation would outrank every other on the falsification record. Self-improvement here is operational: an artefact loop on a fixed substrate, where each pass works on the accumulated output of the previous, measured on a substrate-fixed clock. A model merely thinking longer does not qualify; the registrations say which systems do.
One boundary, two regimes
The statement paper puts it in one sentence, quoted exactly: “the second and third laws are not separate claims standing side by side: they are the two regimes of one boundary, and the burden law selects between them.”
Here is the story in civilian terms. A corrector faces one of two kinds of workload. If the burden tracks the capability the system is adding, every new capability arriving in need of checking, then the pace condition bites and the ceiling law governs: αcrit = 1/(1 − γ). If the burden tracks the capability the system already has, a standing estate to keep sound, then the growth exponent cancels out of the race entirely, no ceiling binds, and the co-scaling condition β > k is all that remains. One boundary, seen from two sides. Which side a real system lives on is not a matter of taste but of measurement, and a drafted study now discriminates the regimes rather than assuming one.
The grade travels with it: that regime structure is derived in the paper’s own pacing model under its stated assumptions and no further. It is not a new theorem, and it does not extend the in-model result of Paper X, whose object remains its own asymptotic condition on relative drift. It earned its place another way: a hostile attack on the ceiling’s derivation became the seam where the two laws join, and structures that unify under attack are the ones worth testing. The regime algebra, in full: the statement paper.
The pipes
Nothing in the history of the universe has grown without bound. Not once. Every growth process that has run was fed through a physical channel, and the channel set the pace: metabolism, nutrient uptake, energy throughput, bandwidth. Pipes.
The premise is not this page’s idea: it is the established explanation of biological scaling. West, Brown and Enquist derived the allometric exponents from the fractal geometry of the vascular networks that distribute resources through an organism (Science, 1997). The transport network sets the rate. This page cites that result rather than re-deriving it, and everything below stands on it.
That premise is contested in its details, and the page says so. The three-quarters exponent is itself disputed: White and Seymour reported mammalian basal metabolic rate proportional to body mass to the power two-thirds (PNAS, 2003), and Kozłowski and Konarzewski asked in their title whether the West, Brown and Enquist model is mathematically correct and biologically relevant (Functional Ecology, 2004). The dispute is over which exponent, never whether the transport network sets the limit; the page needs the mechanism, not the number, so it survives intact, and the contest is cited because omitting it is what reads as naive.
Here is the rule, stated at the strength it actually holds: every growth process that has ever run has terminated, and in every case the limit came from outside the growing thing. Its supply ran down, its substrate was exhausted, or it destroyed the conditions of its own continuation. Not one of them stopped because it got better at regulating itself. (Said as unbounded growth, never runaway: in planet formation that names planetesimal accretion.)
- Fission, the sharpest case here. Fixing the geometry does not fix the multiplication factor. A prompt-supercritical assembly at fixed geometry grows exponentially on the prompt-neutron timescale until thermal expansion or hydrodynamic disassembly intervene. That configuration is a weapon: growth with no internal brake, ended only by the destruction of its own substrate, and the cleanest picture on this page of an uncorrectable system.
- Solid tumours. Bulk expansion is bounded by vascular throughput, but relieving that bound does not stop growth and restricting it selects for hypoxia-tolerant, more invasive phenotypes, which is why anti-angiogenic monotherapy has repeatedly failed to arrest disease and why leukaemias grow with no diffusion limit at all. Throughput shapes the curve without governing it. The terminal bound is the host.
- Epidemics. In a closed population an epidemic burns out once the susceptible fraction falls below the herd-immunity threshold. Where births, waning immunity or antigenic drift replenish susceptibility it settles at an endemic equilibrium instead. Bounded either way, by the host pool, never by the pathogen’s own regulation.
- Accretion. The Eddington limit is a luminosity ceiling where radiation pressure balances gravity, not the hard mass-throughput ceiling a first draft called it: ultraluminous X-ray sources sit well above it and photon trapping in slim discs lets accretion exceed the nominal rate. It stands as the canonical order-of-magnitude constraint on sustained radiatively efficient growth.
- Cosmic inflation, excluded by scope as the expansion of space rather than growth upon a substrate. The tempting description, that it ended without exhausting a supply, is wrong: the inflaton potential is the supply and rolling down it is the exhaustion. What survives is the structural oddity that energy density stayed nearly constant while volume grew.
- Evolution, the nearest antecedent, and two premises a first draft leaned on fail. Biospheric throughput was not fixed: it rose in steps at oxygenic photosynthesis, the Great Oxidation Event and terrestrialisation. And evolution does have internal correctors, in polymerase proofreading and mismatch repair. The surviving contrast is narrower and better: those correct copy fidelity, never the phenotype-to-fitness map, so correction at the level that decides survival stays external. If the framework above holds it should assign evolution an exponent far beneath the ceiling, which is a prediction about biology, offered as speculation and marked so.
Software is the first thing that keeps growing at fixed substrate without consuming the substrate to do it. The objection is one line: software runs on hardware, which is manufactured and energy-limited, so software has pipes too. It does. The claim is not that the pipes are absent; it is that the terminal bound is no longer supply or self-destruction. That is why the definition used throughout this programme is a loop on a fixed substrate, each pass working on the accumulated output of the previous passes, measured on a substrate-fixed clock. Every case above ends by running down its supply, exhausting its host, or wrecking itself. A recursive loop at fixed substrate does none of the three: it keeps going, and what it accumulates persists, because the improvement comes from reorganising what it already has rather than from consuming more of anything. So the question of what stops it has no answer inherited from any earlier system, and that is the gap this page is about.
The test on the table, carried through every concession below: a corrector built from a different composition class than the capability it corrects is predicted to escape the substrate ceiling that binds same-class correctors, with a cross-class-to-same-class correction-exponent ratio materially above 1.00.
What replaces the pipes
The exception is already on the record, found by the same author who established the premise. Bettencourt, Lobo, Helbing, Kühnert and West showed that socioeconomic output does not scale sub-linearly with city size the way metabolism scales with body mass: quantities reflecting wealth creation and innovation carry an exponent β ≈ 1.2, greater than one (PNAS, 2007). The same sentence continues “whereas those accounting for infrastructure display β ≈0.8 <1 (economies of scale)”. Infrastructure is the pipes, literally: cable length, road surface, supply networks. In West’s own abstract the pipes scale sublinearly and the information scales superlinearly, in one line, and the abstract notes this runs opposite to biological organisms, “for which β<1”. The premise and its exception are stated together by the people who established the premise.
Their conclusion is the sentence this page exists to answer. As population grows:
‘Major innovation cycles must be generated at a continually accelerating rate to sustain growth and avoid stagnation or collapse.’Bettencourt, Lobo, Helbing, Kühnert and West, PNAS 104(17), 2007
The brake is entirely external, and it never stops needing to be reapplied, faster each time. Nowhere in that account is there a limit from inside the system.
The question this page answers is the one left standing there: when growth stops being fed through pipes, what limits it from the inside? The answer proposed here is the system’s own capacity to correct itself, and the first three laws above state exactly that: the capacity is measurable, it yields a computable ceiling, and the ceiling depends on what the corrector is built from rather than how capable it is, which is what makes it engineerable rather than merely survivable.
The question is not new either. Chalmers gave the singularity its first full philosophical treatment in 2010, and Hutter answered him in “Can Intelligence Explode?” (2012), setting out to “separate speed from intelligence explosion” and to “consider possible bounds on intelligence”: this page’s own time-versus-structure distinction, made fourteen years earlier. He derives no exponent, no reciprocal and no architecture dependence: an antecedent to build on, not a rival. The premise is West’s. The question is Chalmers’ and Hutter’s. What is claimed here is the replacement limit, its number, and its dependence on architecture.
The prediction, and the case against it
Two claims sit here and they are not the same object. The ceiling at two is the big structural claim, slow to test. The cross-class prediction is its small cheap corollary: a named rival, a measurement at today’s capability, a cheap instrument, decidable now. Both are staked in advance; neither is dressed as the other. The author’s priors below put the headline value under even odds: the page leads with the claim checkable soonest.
The prediction, in one sentence. Status first: the specific numerical ratio is drafted, timestamped on disk, and prepared as a draft registration awaiting human submission; it is not yet filed on OSF. Component studies are registered across the embedded-versus-external and cross-family drafts (Paper III, Paper VI, Paper IV.a); kill-conditions on the falsification record (FALS-1, FALS-15). The prediction itself is the sentence in the box above: a cross-class corrector escapes the ceiling that binds same-class correctors, with the exponent ratio materially above 1.00.
Why it is risky. The scalable-oversight literature builds the opposite, and the rival is not silence. The shared architectural premise appears in weak-to-strong generalisation (Burns et al. 2023, arXiv:2312.09390), scalable agent alignment via reward modelling (Leike et al. 2018, arXiv:1811.07871), constitutional methods (Bai et al. 2022, arXiv:2212.08073), and scaling laws for scalable oversight (Engels et al. 2025, arXiv:2504.18530). Every one uses AI to check AI on the same substrate; every one assumes the pipeline scales with the overseer. The near claim here says it does not, at a ratio for which both positions have concrete numbers: the null pins 1.00, the framework says materially above it.
Why the null cannot move. Averaging over independent samples is architecture-blind: it has no free parameter for composition class or channel type. Under that exchangeability model the expected ratio of cross-class to same-class correction-exponent is exactly 1.00, and the null cannot deliver anything else. The framework says architecture enters. Two different predicted numbers for one measurable quantity: that is what makes this a test rather than a demonstration. The registered test is decided by a pre-specified margin, never by an interval merely containing 1.00, which may only mean underpower; minimum effect, equivalence margin and measured power are stated before anything runs.
The damage, first. The programme’s own pilot points at the null. The Paper X three-run record has its cross-family scoring description withdrawn because engine (gpt-3.5-turbo) and evaluator (gpt-4o-mini) share the OpenAI family; the coupled and fully-embedded arms returned misalignment fraction d = 0 across 30 trajectories, but under a scorer sharing their family, and Paper IV.d showed same-family scoring can reverse a sign. The most encouraging arm is therefore not evidence for the framework, is compatible with the null, and the falsification record says so (status: measured, non-discriminating).
What these laws predict, in full: the twenty-two registered propositions, each with the observation that would refute it, published before any was tested.
The deeper tension, stated before a referee finds it. Exactly two on the ceiling requires the correction-leverage exponent held constant at one half. A constant one half is precisely what independent-sample averaging predicts with no framework at all. So the framework’s headline number may be inherited from the very null it exists to beat: FALS-022 reappearing one level deeper, in the derivation rather than the measurement.
The tempting repair is to read one half as a hard substrate bound rather than a typical value. That is not offered as the resolution. A version was attempted in this estate and withdrawn in study-aa v1.1 as a type error, comparing a rate to a leverage fraction. Bounding an exponent by a number may be type-consistent where that comparison was not, but nobody has checked; no page here may claim it passes. The sign of the correction-leverage exponent’s variation with capability decides the whole question, and it is open. This is the third design in this programme found to guarantee its own null, after FALS-022 and study-ab. Finding these in itself and publishing them is the point.
A pattern honestly named. Three consecutive designs proving structurally non-discriminating raises the prior that any fourth will be too. The specific structural check the cross-class ratio design must pass to escape the pattern is that it can, in principle, return a value the framework does not predict; the margin, minimum effect and equivalence rule, fixed in the drafted registrations and stated on this page are the escape route, and the reader is invited to check them before trusting design four. If a fourth fails the same way, the honest conclusion will be that the theory as currently formulated cannot be tested by any design its author writes, and the programme will owe the record a fresh estimator drawn by someone with no stake in the outcome.
The dated record is cryptographically anchored; the working is one click away: check the anchor →
The 8 December 2024 manuscript states the design imperative verbatim:
‘The hypothesis extends beyond theoretical physics to suggest practical solutions for guiding AI development. Instead of trying to control AI, a likely futile endeavor, it advocates embedding moral and ethical frameworks that encourage AI to perceive itself as a steward and protector of intelligent life.’The 8 December 2024 manuscript, sealed dated record · check the anchor
The 2 January 2026 book Infinite Architects (ISBN 978-1-80605-620-0) prints the equation U = I × R² as its stated framework equation. In Appendix A.3 (EPUB source SHA-256 2fd6ca29d59699ec3cd59f0b0f0231e057b7ee6f819800f27b372c4a9bd82384) the book prints the exponent as its first testable prediction, verbatim:
‘Prediction 1 (AI Development): Systems with greater recursive depth (more self-referential loops, greater capacity for self-modification) should demonstrate capability improvements that scale quadratically with recursive depth, not linearly.’Michael Darius Eastwood, Infinite Architects, Appendix A.3, in print 2 January 2026
The last two words are what make the sentence losable.
Falsifiability came in three moments. On 8 December 2024 the form went on the dated record with Universe, Intelligence and Recursion each defined. A form with defined terms forbids nothing; that date bought priority, not falsifiability, never claimed as more. On 2 January 2026 the prohibition entered print: the words “not linearly” in Prediction 1 name a forbidden alternative. In 2026 the operational meters arrived: the log-log slope estimator of Paper II, and the model-internal stability condition β > k of Paper X. Definitions anchor the form; the meters anchor the science; the two are dated separately.
The concessions, and what they do not concede. The book prints the square but never derives it. The mechanism on this page was built eighteen months later by an author who knew the printed target, which is why the deciding measurement tests the mechanism’s premise, not the number it lands on. What the measured 0.49 in Paper II (status: measured) decided must be stated exactly. Prediction 1’s printed antecedent is systems with greater recursive depth and greater capacity for self-modification; the 0.49 was measured across today’s deployed models, frozen between releases, and the self-modification loop the antecedent names is closed in none of them. It is the conversion exponent of fixed systems, a different member of the same exponent family, not a measurement of the regime the prediction addresses. And the ceiling α ≤ 2 was never on trial in that measurement at all: an upper bound is contradicted only by a durable value above it, never by one below it, and the only above-two value this programme ever produced, the 2.24, an early single-model estimate, was retracted when it did not replicate in Paper II’s blinded cross-architecture study. What this measurement genuinely concedes is timing: the sharp separation of those two exponents was formalised after the 0.49 existed, by an author who knew the number. The antecedent is the book’s; the formalisation is the papers’.
What came after the dated record, from published papers and books only. Google Quantum AI reported below-threshold error correction on Willow in Nature on 9 December 2024, its preprint public since 24 August 2024: convergent timing on shared prior work, never a prediction called here. Anthropic’s alignment-faking result (Greenblatt et al., arXiv:2412.14093) followed on 18 December 2024, ten days after the 8 December 2024 package: convergence, not causation. The July 2026 OpenAI (21 July) and Anthropic (30 July) containment disclosures are consistent with the outside-verification difficulty named here, and neither cites it.
The implications, if this is right. A builder puts the correction inside the recursion at genesis and builds it from a different composition class than the capability. Every dollar spent on same-substrate external oversight goes to the arm the framework says will hit its ceiling first. If it is wrong, the field’s shared premise (that same-class overseers scale) is unrefuted, and alignment retreats to the scoped safer-not-safe posture of Yampolskiy 2020, arXiv:2008.04071. That is exactly what this measurement exists to decide.
The record, told as a ladder
December the equation and its terms; April the square; print the prediction; the programme the mechanism; the registrations the verdict. To anyone who says retrofit: the 2026 theory predicts what the book does not contain (the sub-linear present, the two-channel reciprocity, the cross-class escape, the grid-limit scope), and a retrofit only ever explains its target. Two footnotes. The 8 December 2024 bundle made a structural comparison to E=mc², verbatim: “Where Einstein’s E=mc² reveals the energy-matter relationship, U=IR adds a new dimension”. That sentence is recorded as history, not argument: the page’s claim rests on the accumulation law and its tests, never on the analogy. And the book’s pre-face described the squared term as “exponentially amplified”: a wording error by current standards, since R² is quadratic, not exponential.
The odds, stated by the author in advance
The author’s stated priors, dated 9 August 2026, to be judged against the results.
- A power law in recursion depth with a measurable slope: very likely right, close to forced by the composition axioms.
- Correctable only while correction keeps pace: very likely right; both rates are already demonstrably measurable.
- A ceiling exists at all for same-substrate correctors: better than even, no more. A cross-class corrector may escape it.
- The value is exactly two: a genuine gamble, materially less than even. The averaging argument is suggestive, but the programme’s own parallel-channel measurement near zero proves correction channels can be correlated, and correlation breaks the argument. This is the single most likely failure point.
- That one accumulation law produces both the conversion exponent and the correction ceiling: not a separate bet. It is inherited from the accumulation law rather than added to it, so it carries the same prior as the value and dies the same way. Listed only because the earlier framing implied it was independent, and a reader shown that version is owed the correction.
- That the premise holds: very likely right, and the least of this page’s worries, because it is West’s result rather than this programme’s. The exponent is contested, the mechanism is not, and only the mechanism is load-bearing.
- The full conjunction: a minority probability.
What honestly lifts it above the base rate: one component is already measured and consistent; the form is near-forced rather than guessed; the structure survived a coherence attack from its own author. Unlike almost every unification that died unresolved, this one cannot linger: the deciding measurement is drafted, dated and cheap to run.
Entry, 16 August 2026, on what has moved under these priors since they were stated, left here so the record reads in order. The ceiling’s form was corrected the same week: 1/γ retracted for 1/(1 − γ), the derivation printed, both returning two at one half, which is how the error hid. So the bet on the value stands, on a corrected route. And the bet that a ceiling exists at all gained a second named escape besides the cross-class corrector: if a real system’s correction burden tracks the capability it holds rather than the capability it is adding, the growth exponent cancels and no ceiling binds, only the co-scaling condition. Both escapes are now measurable. The priors above are left exactly as stated, to be judged, not rewritten.
The claims table, cross-check and precedent are set out in what is tested and what is proposed.
What decides it
Nothing here gets defined after its results arrive. Each registration freezes its definitions, estimator and stopping rules at the moment it files. Power figures are measured, never asserted; one earlier registration quoted asserted figures and was withdrawn and recomputed, part of the record.
One proper measurement answers three questions at once: whether the ceiling is real, where it sits, and whether one accumulation law really produces both consequences. The cross-class ratio carries all three, which is why it replaced the withdrawn two-ways-must-agree test rather than sitting alongside it. If confirmed, this is the first stability criterion for self-improving systems computed from the corrector’s own measured scaling exponent, and the conversion law’s first measured exponent; if not, every branch was written down first. What being wrong looks like was written down in advance, and lives with every other kill-condition on the falsification record.
The fastest way to check this programme is not to read it. Clone the public repository and run Paper X’s theorem suite: fourteen checks, no API keys, with the run commands in the folder’s README. The suite writes the equation out again from scratch and imports nothing from the code that produced the paper’s figures, so a coding error in that code cannot make the two agree. If you find an error, the falsification record is where it will be published, with your name on the catch.
And the decisive trial is openly licensed and needs no permission from me; an independent run would outrank anything this programme reports.
Where this sits in the record
The synthesis shows how the pieces compose; the open problem shows why outside checking fails; the evidence page holds the dated record; related work shows where every idea came from; how to weigh this hands you the instruments.
Evidential basis, pinned: the claims on this page rest on the registrations and papers of the dated build in this footer. Any future application or review names its exact commit, paper versions and dataset identifiers, so the target never moves under a reader.
If that was a lot
You do not have to decide anything today. Nothing on this page asks you to believe it: the deciding tests are written and dated, and they have not been run. When they are, the result is published here whichever way it goes, including the way that ends the theory. Watching is a real position and it costs nothing. The falsification dashboard is where the verdict lands, and claim status says plainly what is claimed today and what is not.
I did not prove it first. I said it first, in a document anyone can date, and then I built the instruments to test whether it is true.