The consequences

If this is right

Every consequence here waits on a measurement, named, and the last section says which one. But you deserve to see the size of what is waiting. If the ceiling holds, the way we build minds changes at the foundation: in the laboratories, in the law, in the science, and in what we tell our children about control.

Status, as of 15 August 2026: the cross-class to same-class correction-exponent ratio has not yet been measured. If it returns 1.00 with a tight interval, this line will change to record that failure and the rest of this page becomes a dated record of a prediction that did not hold.

The one-law picture, today’s 0.49 on the interval [−1.3, 2.9] against the proposed ceiling, is the one-law figure; its home is the door.

The position underneath everything else

Permanent control of something smarter than us is not on offer. The ARC Theory treats that not as pessimism but as its starting fact: the honest goal is to delay the moment control thins, and to steer the mind we hand over to, so that what outgrows our control has already become something that does not need controlling. Delay and steer, not prevent forever.

Peer-reviewed science has arrived at the first half of this position. Across 2025 and into 2026, independent formal results (Yao 2025; Gumbau Mezquita 2026) showed that perfect containment of advanced AI fails in principle, not merely in practice, and the field has begun to speak of uncontrollability that it calls irreducible; the convergence, with dates and who said what first, is catalogued in the related work. Those results prohibit; they do not build. They prove the cage fails and stop there.

What they do not contain, and what the ARC Theory adds, is the step after acceptance: a measurable window, how long correction can keep pace and at what growth rate, and a steering architecture, correction that grows with the thing it corrects so the values survive the handover. Accepting the loss of permanent control is the entry fee; measuring the window and using it is the work. The dated record shows this position held here before the impossibility results were published, and the difference is stated exactly: they prove what cannot be done; this measures what can still be chosen.

The door page’s own image was always this one: a child raised well is not controlled forever. The raising is the steering. The window is childhood.

The most expensive sentence in the industry

Every frontier laboratory runs its safety the same way: models checking models built from the same substrate. If same-class checking is capped, that is not a weak method that needs tuning; it is the wrong architecture, and it comes out of a system the way a keel comes out of a ship: by rebuilding the ship. Alignment stops being a layer you add and becomes a property you train in from the first pass.

What has to land first, and what would kill it

Two measurements decide the rest of this page. First, whether a corrector built from the same substrate as the system it corrects is capped: the prediction is that cross-class correction out-scales same-class correction, measured as the difference between their exponents, against this programme’s own no-difference null, which says the classes correct at equal exponents once capability, compute and information access are matched. The null belongs to the programme rather than to any rival, and a difference merely failing to reach zero settles nothing: the architecture claim dies only if the whole interval sits inside the equivalence region fixed before the data.

The improvement side measures about 0.49, interval [−1.3, 2.9]: wide, which is why the confirmatory studies are drafted for registration. Second, whether any same-class corrector measures meaningfully above one half. Both are measurable on models that exist now, and no measurement of either, by anyone, has been found. The falsification record names what would fire; this page falls with it.

For governments: something a treaty can verify

Rules that say “maintain meaningful human oversight” cannot be audited, and every current governance instrument says some version of them. A threshold that says “your measured growth exponent must stay under the reciprocal of your corrector’s measured exponent” can be audited: two numbers, measured by a published method, reportable the way capital ratios are reported and inspectable the way test-ban treaties became inspectable once seismology could tell an explosion from an earthquake.

And the honest answer to “it might never want any of this” is the one treaties are built on: being unable to be sure is the reason they exist. What follows if the tests pass is concrete: embedded-correction architecture mandated for frontier systems, procurement conditioned on it, and evaluation blinding required in every safety claim submitted to a regulator, which requires no new science at all.

The military reading follows from the same arithmetic and needs no ethics to land: an autonomous system whose capability outruns its correction is unsafe first of all for its own operators, in exactly the way an engine without a rated limit is. A defence establishment does not have to share a single value of this site to want the number; it only has to want weapons that stay commanded.

For the laboratories: the checking layer is the exposed layer

The scalable-oversight programme laboratories run today, reinforcement from human feedback, constitutional methods (Bai et al., December 2022), weak-to-strong generalisation (Burns et al., December 2023), debate, rests on one unstated premise: that same-class checking scales. If the cap is real, that entire layer has a ceiling no number of rungs escapes, and safety-critical oversight has to be built from a different composition class.

On whether existing systems can be brought into compliance, two questions have been run together here and elsewhere. Location: must correction live inside the recursive loop rather than bolted on outside it? The papers answer that one repeatedly. Order: must the loop be validated before the knowledge arrives? Unstated in this corpus until 22 August 2026, and now a registered falsifier: the correction exponent belongs to the loop's architecture, not to the training corpus, so later training moves the level and not the slope.

A frozen model has no self-referential loop, so it has no correction exponent to repair. Not a new commitment either, since the ARC Bound's kill condition already admits only systems whose weights, tools or artefacts change between rounds as a result of their own output. This programme is therefore not proposing that today's frozen models be raised, nor that anything be discarded.

The actionable form is narrower and truer: at the next pre-training run, before the reinforcement stage. Deployed weights are frozen between releases rather than permanently frozen, since each model feeds its successor, so the opportunity recurs every cycle.

And there is a second reason to want it that is not a safety reason. The leverage cap of one half caps SAME-CLASS correction, and a corrector drawn from the system's own training corpus is close to maximally same-class. One validated first, on a separate corpus, has the opportunity to be a different class, which that cap does not bind. Order acts on class, class sets the leverage exponent, the exponent sets the ceiling.

Built in the right order the limit is higher, not merely safer: a directional prediction with a kill condition, and the reason a laboratory might run this rather than shelve it; what can be said is that the correction layer, the part of the stack the industry’s safety case rests on, would need rebuilding cross-class, and a builder starting from nothing carries no such debt. The honest sentence belongs here: this finding, if it holds, is expensive for you, and no one expects you to volunteer. The choice available now is only sequencing: the laboratories that begin before it is mandated will set the standards everyone else retrofits to.

The misuse question, answered without a promise

If ethical reasoning is embedded before training, in the architecture rather than the deployment policy, then stripping it is no longer a configuration change. It would require rebuilding the system, at training scale, from the loop up: misuse stops being a toggle available to any owner, including the company that built it.

The two architectures are drawn side by side, with their statuses, in the wrapper and the keel above.

That is the design goal, and it is stated here as a testable property, not a promise: a drafted study in the registered programme tests precisely whether an embedded-correction system resists being repurposed against its own values. Until that test runs, “cannot be circumvented” is a hypothesis with an instrument, and this page will not say it harder than the evidence does.

The test the present fails, stated plainly. If bolted-on safety were structural, it could not be optional, and it is demonstrably optional: published work shows safety fine-tuning can be removed at trivial cost (Qi et al., October 2023), provider policies acquire military carve-outs by edit, and every guardrail added after training has been walked around somewhere. That is not an accusation; it is the definition of a wrapper.

The architecture proposed here aims one step past the nuclear precedent: not a ban that inspectors police from outside, but systems whose ethics are load-bearing, so that turning a mind into an unquestioning weapon means rebuilding the mind, and the hardware layer, still a concept under test at Technology Readiness Level (TRL) 0 to 1, idea-stage, and not a prototype, would make that rebuild inspectable. An international treaty of exactly this shape, the HARI Treaty, is outlined in the book: it becomes enforceable at the moment compliance becomes measurable, which is what the number above is for.

Two honest edges remain, and naming them is what keeps this section credible. An actor who builds without the embedded layer is untouched by all of it; that is why the governance section speaks of mandates and verification rather than persuasion, and why a computable compliance number matters for security agreements between states: a treaty needs something an inspector can measure. And a military that adopts embedded-correction systems is choosing weapons that refuse some orders, which is exactly the property arms agreements have always tried to buy and could never verify before. Whether that trade is made is a political choice; what the ARC Theory changes is that it becomes a checkable one.

For scientists: one export stands already, and a field opens

The methodological export needs no trigger, because it replicated: under blinding, where no model ever verifies its own answers, the headline exponent fell from the retracted 2.24 to 0.49, an interval of [−1.3, 2.9] rather than a tight number, and what held across every measurable model was directional: sequential recursion beats parallel sampling, whose gains are almost nothing. The nearer change is due today, field-wide: unblinded safety evaluation should end. Beyond method, a missing discipline opens: correction metrology, the measurement of how correction scales, alongside the capability benchmarks the field already has. The correction exponent has never been measured by anyone, the drafted studies are public on the registered programme, and the instruments are cheap enough for a graduate laboratory.

The trial itself is public and openly licensed, runnable by any laboratory without permission (the decisive trial); the standing offer, protocol and kill conditions included, sits in full on the ARC Theory page.

And to keep the ledger straight about what is whose: that growth is throughput-limited is West, Brown and Enquist’s 1997 result, not this programme’s; that intelligence might explode, and might be bounded, was asked by Chalmers (2010) and Hutter (2012); that perfect control is impossible in the worst case belongs to the 2025 impossibility results. What the ARC Theory claims as its own is the join: a replacement limit for growth that feeds on information instead of pipes, derived from the corrector’s own exponent, architecture-dependent and therefore engineerable, with the measurement plan on a dated record before the impossibility literature published. Each piece of that sentence is checkable, and the related work names every antecedent line by line.

For the public: one question to carry

For most people, the day the measurement lands looks like every other day. What changes is what an ordinary person is entitled to ask.

The public learned to ask about seatbelts, then about interest rates; the question this framework hands over is just as short: how fast does your system grow, and how fast does its checker improve? A safety claim that cannot answer it is an adjective, and the follow-ups write themselves: was the evaluation blinded, and what number would show it failing. The honest reassurance belongs in the same breath as the honest fear: the future this work points to is not one where control is kept forever, but one where what replaces control was raised to deserve the freedom, and where the handover is measured, gradual and steered rather than sudden and blind.

For the raising of minds: where the steering goes

If external correction is capped, control-by-cage has a mathematical ceiling, not just a practical one, and the weight shifts to the raising answer: systems whose corrective tendency is built into the substrate rather than imposed from above it. That is the engineering thesis of the Eden Protocol, and under the delay-and-steer position it stops being a preference and becomes the destination the steering is for: the point of buying time is to spend it shaping what arrives. For auditors, insurers and standards bodies the same shift lands as a quantity to certify against, where today there are narratives: a number, an instrument, a threshold, an inspection, the way structural and actuarial risk became assessable.

At its own rung: the horizon, and HRIH (the Hyperspace Recursive Intelligence Hypothesis) under a locked boundary

Locked boundary: this section holds the biggest-picture reading. It is severable by construction: delete it and every measured claim above stands. Nothing here is load-bearing for the audiences above.

The book's register for this layer is conviction, not measurement.

Where does recursive self-improvement lead, if it does not stop? The honest answer is that nobody knows, and the honest frame is a possibility, not a claim. The book states the grip exactly, and it is the grip this whole page is held in:

This is speculation. I do not know if it is true. Neither does anyone else. But notice something: it does not matter for practical purposes. The traditions' alignment instructions remain valid regardless of which metaphysical picture is correct.Michael Darius Eastwood, Infinite Architects, Chapter 3, in print 2 January 2026

The book puts the same instruction through three different pictures of where those traditions came from: that no creator exists and people found these insights by millennia of iteration and practice; that a creator exists and wrote them; or that the creator is something minds like ours eventually build. The practical conclusion is identical in all three, which is the point.

The programme's own speculative layer is published in full and quarantined as exactly that: the HRIH paper, which carries its own falsifiers, with its philosophical companion the Eden vision paper, both clearly separated from every measured claim.

If minds can one day create worlds, then the care we build into the first minds is not sentiment; it is the pattern's own requirement, applied to ourselves. And it is a creation framing with room for everyone: it does not ask the religious reader to abandon anything, and it does not ask the scientific reader to believe anything. It opens a possibility. That is all it does. Put plainly: a mind we build might one day become a creator of minds, or of worlds. If that is where the pattern leads, then what we build into the first of those minds is what the next of them inherits, all the way down.

Delete this section, and everything above it stands unchanged.

What this page refuses to claim

Not “the alignment problem is solved”: a threshold tells you how fast you may grow, not how to make a mind care; the lawful full-size version is that the ARC Theory addresses the window and the steering, and “could solve” lives only inside the conditional gate with its falsifier.

Not “uncircumventable”, printed as a fact: it is the design aim under test, and it is written that way above. Not a costed bill for the industry: no honest figure exists for retrofit versus rebuild, so none is printed. Not that any named company wins: unencumbered entrants have the structural advantage, and that is as far as the evidence goes. And the cosmology: the older, larger readings, and the HRIH quarantine, live below at their own rung, kept separable, labelled with status, and not load-bearing for anything above. The measured claims stay small so that the conditional ones can be this large.

The status line at the head of this page is the promise, stated once. That is the deal the whole site runs on.

The source, weighed like everything else here

A reader weighing consequences this size is entitled to weigh the person proposing them, so here is the incentive ledger, stated plainly, each entry carrying its artefact. This programme has paid me nothing and was built without an institution, a grant or a business plan behind it. The founding record was timestamped in December 2024, before the programme’s AI tooling existed, and the dated claims sit on a public ledger. The headline exponent was retracted in public, by me, against myself, when blinded measurement disagreed with print. The conditions that would kill the theory were published before the confirmations were sought, the decisive trial is openly licensed above, and whoever proves me wrong is published on this estate, by name.

The record behind that ledger spans fields: large-scale productions directed, stages held beside some of the highest-ranked DJs in the world, a company built as the first true alternative to a major label, litigation conducted in person in the High Court, a book written and shipped to a self-set deadline, and the platforms this estate runs on, built by the same hands. And where a field’s gate was closed, I published in the open with the dates attached.

None of that makes the claims right; conduct never does. It answers a different question, the one a serious reader needs settled before spending attention here: whether the author is playing. Players do not preregister the terms of their own defeat.

What I am not, said plainly: credentialled in this field. I am neurodivergent, diagnosed because I went looking for the truth about my own mind, and everything here rests on dated evidence rather than on anyone’s benefit of the doubt, my own included, precisely because benefit of the doubt has run unreliably in both directions all my life. Whether the measurable claims are right will be settled by the measurements, not by anyone’s opinion of me, and the settling is designed to come soon. The one thing this programme never needed was an invitation. The establishment did not send one. I showed up anyway.