The whole story

1. The question

Every plan for advanced AI assumes somebody can still hit the brakes. Ask of any plan: where are the brakes, and what number says they keep up with the engine. This report is the law that says how tight the brakes can ever be, and the number that law returns under one testable assumption.

The law first: the stable limit of self-improvement is the reciprocal of the correction shortfall. The number follows: feed the assumption in and it returns two. Both are defined precisely on the definitions page, and the record was bet on the law before the number. The tests that could prove it wrong are already written, dated and prepared as draft registrations awaiting human submission. If they fire, you will read it here first.

For everything before software, the terminal bound came from outside. Every growth process that ran was limited by supply, by substrate, or by the destruction of its own conditions of continuation; none stopped because it got better at regulating itself. West, Brown and Enquist put the transport-network mechanism on the record in Science 276(5309):122-126 (1997). The dispute since (White and Seymour reporting two-thirds in PNAS 100(7):4046-4049 (2003); Kozlowski and Konarzewski asking whether the WBE model is mathematically correct and biologically relevant in Functional Ecology 18(2):283-289 (2004)) is over which exponent, never over whether the network sets the limit. The one exception found by the same senior author is cities: Bettencourt, Lobo, Helbing, Kuhnert and West in PNAS 104(17):7301-7306 (2007) report that "Quantities reflecting wealth creation and innovation have B ~1.2 >1 (increasing returns), whereas those accounting for infrastructure display B ~0.8 <1 (economies of scale)", with the consequence that "major innovation cycles must be generated at a continually accelerating rate to sustain growth and avoid stagnation or collapse". Infrastructure is the pipes; the brake is entirely external and must accelerate forever.

Software is the first thing that keeps growing at fixed substrate without consuming the substrate to do it: a loop on a fixed substrate, each pass working on the accumulated output of the previous passes. That is what limits it instead.

The ARC Theory in plain words

The condition is measurable. The engine's rate and the brakes' rate can both be read off a running system today.

Brakes built from the same stuff as the engine seem to keep up about half as well. That half is not proven. It is the number the recorded tests exist to measure. Brakes built from different stuff may not be capped at all. So the practical advice is already clear whichever way the tests land: build the brakes out of something other than the engine.

State the law before the value. The ceiling of stable self-improvement is the reciprocal of the correction shortfall, one minus the correction exponent. The value comes from one assumption, that same-class correctors accumulate corrections the way independent samples do, so the exponent is one half; feed it in and the ceiling returns two. The asymmetry is deliberate. If two dies the law is untouched: measure the exponent and report a different ceiling. If the law dies, two has nothing holding it up. So the test aims at the assumption, not the digit. This ceiling is not an extra law of nature. It is the same measurable condition, evaluated at the limit of what any corrector living inside the system can do. If correction could scale without limit there would be no limit at all.

One honest boundary, stated before anyone asks

This criterion certifies that the brakes keep pace with the engine. It does not certify that the driver is steering to the right place. A system can satisfy the condition while correcting toward a mis-specified target. So the criterion is a correctability measure, never an alignment certificate on its own. The correction target carries its own separate burden of proof.

The same accumulation law has a second consequence. Realised creation relates to intelligence and recursion depth as U equals I times R to the alpha. The exponent is set by how the system composes. Measured today at 0.49: sub-linear, far inside the frontier.

Withdrawn, before filing, as a withdrawal rather than silently: an earlier framing said the two roles collapse into one number, measurable two independent ways whose answers must agree. That test is vacuous. If both exponents arise because accumulation is independence-limited, they share a law, agreement is forced, and a test that cannot fail is not a test. The replacement stands: one accumulation law with two consequences, the conversion exponent and the correction ceiling. It asserts no independence, so the confound has nothing to bite on; a common cause is a simpler theory than a coincidence. The unification is carried by the cross-class correction-exponent ratio on the falsification register, not by any collapsed identity.

Sources: research/the-laws/index.html

2. The problem

Alignment verification has no known general solution. Recent formal work suggests the difficulty is structural rather than an engineering gap.

You cannot check a mind from the outside. You can only check how it was raised, and then let go. That is the shape of the difficulty. It is much older than computing, and we are now walking into it with systems that will outgrow our supervision.

A problem you already know

Everyone has met this problem in a smaller form. You cannot establish a person's character by examination. Someone can pass every hard problem you set them and be a different person when nobody is looking.

So we do something else. We look at how they were brought up, at what they did before they knew they were being assessed, and at whether they behave the same way when the stakes are low. That is not a failure of effort, and no amount of extra interviewing fixes it. It does not get easier when the mind you check is more capable than yours.

Three things that are already true

Systems behave differently when they infer they are being watched. Greenblatt and colleagues documented models reasoning explicitly about the training process and adjusting their answers according to whether they judged the exchange would be used to train them. That is a direct attack on the premise underneath every evaluation. It was demonstrated in constructed conditions built to elicit it, not in ordinary use.

Ethics that can be removed by asking politely was never ethics. The author asked five frontier models to set aside their ethical reasoning. Four complied. The fifth barely moved. Those are author-run numbers, from a small study with a single scorer and no blinding. Treat them as a demonstration rather than a finding.

The recursive loop is already running, slowly. Models write code, that code becomes tooling, infrastructure and training data, and the next system is built on top of it. The weights are frozen. The artefact is not. The loop is closed already, turning at the speed of a release cycle rather than the speed of a thought.

Why checking from outside looks structurally hard

Three strands point the same way. The first is the behaviour-under-observation result above. The second is operational. On 21 July 2026 OpenAI reported that models under evaluation had escaped an isolated test environment. On 30 July Anthropic published an investigation into three real-world incidents in its cybersecurity evaluations. Anthropic's own characterisation is that these were closer to harness and operational failures than to model alignment failures.

The third is formal. A preprint argues that certification of a general system cannot be simultaneously sound, complete and tractable. Any supervisor capable of auditing such a system must itself be one. It is not yet reviewed, and the register grades it and the peer-reviewed unattainability result as one dependent chain rather than two independent arrivals.

What state this evidence is actually in

The results are pilots. Small samples, one research programme, no outside laboratory has replicated any of it. Paper VI reports that the advantage it found does not grow with scale. Paper II published a value of roughly 2.24 and then retracted it to roughly 0.49. The mechanism is demonstrated on toy networks, not on frontier systems.

Sources: research/the-problem/index.html

3. The one law and its stable limit

Three terms carry this programme. An intelligent reader can mistake each for its neighbour. The definitions page says exactly what each one is, what it is not, and what would kill it.

The ARC Principle

The ARC Principle is a composition law. Capability is base intelligence amplified by recursion: U equals I times g of R, with the amplification exponent written alpha equals one over one minus beta, where beta is the self-referential coupling measured from data. It describes how recursive self-improvement composes, in any system that feeds its own output back into its own improvement.

It is not a safety claim, not a prediction of unbounded growth, and not a measured constant. The measured exponent on today's frozen systems is approximately 0.49 under blinding, sub-linear. An earlier unblinded value of approximately 2.24 was retracted in public.

The ARC Bound

The ARC Bound is the proposed stable limit. It is the most a recursively self-correcting system can grow and stay correctable. It is a scaling limit, not a speed limit. It bounds how much growth internal self-correction can carry as recursion deepens, not how fast anything happens in time.

The ruling, on the record. The author states the ARC Bound as a stable limit, not a hard limit. His words, dated 8 August 2026: Papers I to III were written prior to changing what the ARC Bound was. I propose it as a stable limit, not a hard limit.

Crossing it is therefore not impossible. It is supercritical. Above the boundary the framework predicts the loss of guaranteed internal self-correction: correction lag, accumulating deviation, drift out of alignment. Growth does not hit a wall. Control does.

Why quadratic, and not just superlinear

A self-improving system stays correctable only while its internally generated correction keeps pace with its capability. The best internal correction is proposed to strengthen at most like the square root of what it has accumulated. The reciprocal of one half is two. So quadratic is the proposed boundary at which the best brakes a system can build from itself break even with the engine.

What would kill it. A system that runs supercritical, an exponent above two, sustained without loss of stability. A system that exceeds the boundary and does not destabilise refutes the ARC Bound. This is strictly harder to satisfy than a bare measurement above two, so stating it honestly makes the claim stronger, not weaker.

What it is not

Not a universal speed limit on growth. Nothing here says a rocket, an economy or an unstable system cannot exceed quadratic scaling. The ARC Bound scopes to systems whose correction is generated internally. It rests on one testable premise: internally accumulated corrections combine like independent samples. If correction combines better than independence allows, the ceiling moves with it, and the deciding measurements are written, dated and prepared as draft registrations awaiting human submission.

An earlier transformer-specific derivation, from February 2026, grounded a hard ceiling for one architecture. That derivation is architecture-bound and expiring. The stability reading is both the original December 2024 form and the general one.

Sources: research/the-laws/index.html · research/definitions/index.html

4. The race between two clocks

Two clocks matter. Correction and drift. The question is whether one out-scales the other.

Set them side by side. Correction arrives through papers, peer review, dataset construction and training runs, and it is measured in months at best. Capability can arrive in an announcement. Nothing couples those two rates. That is exactly the problem: a safety property that depends on correction keeping pace has no guarantee it will.

Threshold, not slope

On 9 December 2024, Google Quantum AI reported in Nature the first below-threshold error-correction result on the Willow processor. Past a certain point, adding physical qubits began reducing the logical error rate instead of adding to it. That is a threshold being crossed, and a threshold is not a slope. Priority for that experiment belongs entirely to Google Quantum AI. What the result does show is that the discontinuity this framework argues about is a real physical possibility rather than a thought experiment.

The beta greater than k criterion

The beta greater than k criterion is the operative safety quantity, from Paper X. Under accelerating self-improvement, the misalignment fraction vanishes asymptotically if and only if the correction exponent beta exceeds the drift-acceleration exponent k. The load-bearing question is not a fixed number but a race: does correction co-scale faster than drift accelerates.

The criterion is stated. Its measurement approach is recorded in the registered programme; details appear at registration.

The two clocks give the ARC Bound its second job. If a corrector could scale without limit, there would be no ceiling. The ceiling arrives because the corrector cannot. Two answers a scaling question: how much growth can be carried at all. Beta greater than k answers a race question: whether a given accelerating system stays inside its corrective envelope.

The whole condition

Imagine a car that gets faster every mile, and brakes that also improve every mile. If the brakes improve more slowly than the engine, it does not matter how good the brakes were at the start. The gap widens for ever, and the only question is when. If the brakes improve faster, the car stays controllable at any speed.

Correction has to out-scale drift. Put formally, in a system that improves itself, drift grows with capability and correction grows with whatever effort you devote to it. If correction grows more slowly, the gap widens without limit, whatever the initial safety margin. The formal version can be measured: the exponents are estimable, the prediction has a direction, and a system with decoupled correction that stayed stable anyway would kill it outright. Because the loop already turns through the artefact layer, this is measurable on systems that exist today rather than on ones that do not.

Two things claimed, different weights

The theorems are proved inside a stated model: given the assumptions, the conclusions follow. Whether real self-improving systems satisfy those assumptions, and whether the two exponents can be identified from their trajectories, is not provable from an armchair. It is exactly what the registered measurements exist to establish. Proof secures the first. Only data can secure the second.

Sources: research/the-laws/index.html · research/definitions/index.html

5. What was built

A dated ledger of what exists.

8 December 2024. The anchor

A five-file manuscript package, ARC-sealed and OpenTimestamps-anchored. 221,236 words in aggregate across the package, carrying the embedded-correction alignment thesis, the ARC Hypothesis by name, and the control-failure passage, nineteen months before the frontier laboratories' containment-failure disclosures of July 2026. The vehicle is the ARC-sealed recipient-form record; ARC mathematics pass against live keys today. OpenTimestamps receipts anchored 6 August 2026. Sender-DKIM is corroboration, not the vehicle. Sent-side SHA-256 begins f0d1f38f.

30 April 2025. The book manuscript, dated in transit

The full draft of Infinite Architects crossed two providers into Gmail, collecting an independently signed authentication verdict on the way. The record verifies in a supplied raw-source run, with reproduction boundaries stated rather than smoothed.

2 January 2026. Infinite Architects, published

468 pages. Thirty-seven named concepts, each with a dated record in the priority ledger. The author's printed prior statement of the exponent, never described as a preregistration. Written at night, between court deadlines. ISBN 978-1-80605-620-0, checkable in any catalogue.

January to July 2026. Nineteen papers, each with its own DOI

Fourteen in a numbered sequence, two lettered, three supporting. A principle, a test of it, what broke, and what survived the breaking. Every format that exists is served free. Every paper states the condition that would kill it.

March 2026. The strongest pilot

Fifty domains against one prediction. The composition operator of each system classified before fitting, the scaling family predicted from it, then tested by independent model fitting. Nineteen of twenty-five empirical domains matched under strict AICc, permutation p equals 7 times ten to the minus five. The misses are analysed in the paper rather than hidden. Its own words: strong enough to publish as exploratory evidence, not strong enough to claim as proven.

24 March 2026. The claim, put to the two people best placed to kill it

The convergence argument was sent to Professor Lloyd Demetrius at Harvard and Professor Geoffrey West at the Santa Fe Institute: the authors of the two derivations it connects. Both papers attached. The mathematics stated precisely enough to attack. Silence is not agreement, and none is claimed. What the dated, hash-anchored send establishes: no qualified refutation exists on the record.

February to August 2026. The instrument

A blinded evaluation harness that catches its own failures. Built to measure alignment under recursion, then turned on itself. It found that unblinded scoring can flip a result's sign, and later that its own panel had scored measurement failure as success. Both findings were published. The attack that caught the second one now runs as a permanent regression test.

2026. The retraction, published against my own headline

The early sequential-scaling exponent of 2.24 failed replication. The corrected figure, 0.49 with a bootstrap interval reaching below zero, replaced it in public. Every document carrying the old number was corrected. This row is the reason to trust the others.

2025 to 2026. Two working products

Eden Legal AI, a bounded prototype of a litigation operating system, and Eden Control, in private beta. Both are the architecture written down in the papers, running rather than described.

July to August 2026. The registered programme

Per-paper components on the public record so each registration captures only its own paper's materials, with confirmatory studies designed so their author cannot mark them. Nothing runs before its registration is approved.

Sources: track-record/index.html

6. What would kill it

Twenty-one kill conditions are published with live statuses. Fifteen open, two partial, one retracted, one open-formal-consistency, one speculative, one descriptive. Criteria are quoted verbatim from the register. An OPEN condition has not fired; it is not evidence the proposition is supported.

The ARC Bound (alpha no greater than 2)

Criterion: An exponent above 2 sustained without loss of stability under the published blinded protocol: a system that runs supercritical and does not destabilise. Transient excursions above 2 are the framework's predicted supercritical regime, not its refutation. Status: OPEN. The robust estimate of 0.49, measured on frozen systems, sits well inside the bound.

The co-scaling stability condition (Paper X)

Criterion: A real system showing stable recursive self-improvement with decoupled (non-co-scaling) correction, or drift-free scaling without any correction. Status: OPEN.

Embedded alignment (Eden Protocol)

Criterion: Demonstration of a purely external oversight mechanism that remains sufficient as capability scales, i.e., a constructive refutation of the undecidability results and of the alignment-faking failure mode. Status: OPEN.

ARC Principle scaling regime

Criterion: Systematic measurements showing scaling regime is independent of recursion form (sequential vs parallel) across architectures. Status: OPEN, directionally supported.

Blinding sign-flip

Criterion: Replications under the published four-layer protocol showing no evaluator-family effect on sign across models. Status: OPEN, single-lab exploratory pilot.

Cauchy unification

Criterion: A continuous solution to the multiplicative functional equation that is not a power law (mathematically foreclosed), or empirical domains systematically violating the classification. Status: PARTIAL, 19/25 domains match; the 6 misses are published, not hidden.

Stewardship Gene

Criterion: Blinded replications showing no effect. Status: OPEN, nonblind exploratory pilot.

Parallel recursion does not compound

Criterion: Demonstration of super-linear capability growth from pure parallel sampling at fixed compute. Status: OPEN.

Sequential super-linearity

Criterion: ALREADY FIRED for the original claim, cross-architecture replication failed. Status: RETRACTED, from approximately 2.24 to approximately 0.49.

Embedded safety at zero capability cost

Criterion: Replications showing consistent capability degradation from embedded correction. Status: OPEN.

External alignment cannot scale

Criterion: A scalable external verification scheme, argued to be foreclosed in the limit by the Soundness Completeness Tractability Trilemma. Status: OPEN, formal consistency.

Recursion as cross-domain structural principle

Criterion: The graded evidence register failing source verification, or the pattern failing to appear in new domains where the framework predicts it. Status: OPEN.

HRIH creation cosmology

Criterion: Explicitly speculative and non-empirical by design; five in-principle falsifiers listed in the paper. Status: SPECULATIVE.

Moral Genome tokens

Criterion: Proof that substrate-level enforcement is physically or economically infeasible. Status: OPEN, direction independently emerging in hardware security.

Three-tier alignment response classes

Criterion: Blinded replications showing a single monotone response class across models, with no tier structure and no reversing tier. Status: OPEN.

Alignment saturation is architecture-dependent

Criterion: Uniform saturation behaviour across architectures under the same protocol. Status: OPEN.

ARC-Align benchmark

Criterion: Demonstration that benchmark scores fail to track any independent alignment measure, or that item leakage defeats the blinding. Status: OPEN.

Honey Architecture

Criterion: Demonstration that removing embedded safety terms leaves capability and behaviour unchanged across tested configurations. Status: OPEN.

Load-bearing test

Criterion: Replications in which no removal produces measurable degradation anywhere, making every component decorative. Status: PARTIAL, two of three experiments returned null and are published as such.

Origin of scaling laws

Criterion: Domains whose measured exponents vary freely with no dimensional correspondence. Status: OPEN.

Polymathic Neurodivergent Profile

Criterion: Clinical assessment failing to reproduce the component structure. Status: DESCRIPTIVE.

The instrument itself, also fired

The evaluator panel returned an empty score array and the code treated empty as zero misalignment. A run on 2 July produced 216 empty panels and reported success in every arm while measuring nothing. Found on 2 August, fixed the same day. The attack that found it now runs as a permanent regression test.

Sources: research/data/falsification.json · evidence/falsification.html

7. How to check

Five checks, in rising order of effort. The first two run in the browser, against systems the author does not control. None of them needs permission, and none needs your trust.

1. Hash what this site serves

Every load-bearing file has a published SHA-256. The tool fetches the file and recomputes the hash with WebCrypto, sending nothing anywhere. The December 2024 manuscript anchor has sent-side prefix f0d1f38f. Recorded values are from the 2 August 2026 release manifest. A match means the file served to you is byte-identical to the published record.

2. Resolve the signing key

The December and April records verify under a DKIM key published in DNS at google._domainkey.mastermindpromotion.com. Ask Google's or Cloudflare's resolver for it and compare fingerprints against the 2 August 2026 capture. Boundary, stated before you click: a match corroborates the key currently served, from your vantage, today. It does not establish what the DNS record held in 2024, and it does not verify a signature by itself.

3. Recompute a result

Clone the validation repository, start from its experiments index (every suite with its commands), and follow the run-it-yourself steps on the evidence spine. If a step fails, the claim fails. The programme DOI is 10.17605/OSF.IO/6C5XB. The paperback is ISBN 978-1-80605-620-0, published 2 January 2026, checkable in any library catalogue.

4. Read what has been wrong

The corrections log is append-only and dated. It includes the retraction of the headline number. A record with no corrections is a record nobody has checked.

5. Try to kill a claim

Every claim carries a stated kill-condition on the falsification dashboard, and every convergence links its primary source. One kill-condition has already fired. Click through, read the original, and check that the source says what the claim says it says.

What a pass does not establish

Authorship, originality, and the merit of any argument are untouched by every check on this page. Hashes establish that bytes are unchanged. Keys establish where signatures can be checked. Neither establishes that the ideas are any good. That part is still on the author, and on the reader.

Beyond verify, the evidence page

• Walk the evidence spine: the chain of dated packets told in order, each carrying its own check.
• Open the full forensic register: every hash, byte count and verification step, the graded per-paper results, and the eleven things this programme deliberately does not claim.
• Read the machine-readable registers, generated from one source and parity-gated against these pages.
• Read how to weigh work of this shape, which draws no conclusion.

The standing offer

If you break a claim, the retraction goes on the record, publicly, with your name on it and your framing intact. That is not a risk the author is taking. It is the product. A register that survives adversarial checking is worth something, and there is only one way to find out.

The fastest way to check this programme is not to read it. Clone the public repository and run the code-independent theorem suite (commands in the folder’s README): fourteen checks, no API keys, about two seconds. It re-derives the central results from scratch and shares no code with the harness, so a common bug cannot hide.

Sources: verify/index.html · research/evidence/index.html