The open problem

Related work and contribution boundary

The synthesis is the claim, and the book said so first, in print, 2 January 2026, before any review ran.

“I make no claim to have invented the components. Recursion is well understood in mathematics. The alignment problem has been articulated by minds far more credentialed than mine. Religious traditions have explored stewardship for millennia. What I offer is the synthesis.” (Author’s Note, 2 January 2026.)

I then ran an adversarial precedence review across ten frontier configurations from six vendors, each instructed to attack. It found the components, as the book had already said. It did not find the arrangement: no single prior work joins them, and not one of the five bodies of prior work it examined contains the scaling relation. Its verdict named the operationalisation as the strongest surviving claim. This page records both halves without softening either.

The review has since gained a second chapter, aimed at the mathematics rather than the literature. In August 2026 two further adversarial audits, commissioned by the author and run externally against the statement paper itself, converged on a real defect: the ceiling law’s earlier printed form, 1/γ, ran the wrong way and had no derivation behind it. The correction is published in the paper with the three-line derivation beside it, the corrected form α_crit = 1/(1−γ) passes every check the old one failed, and the audits’ surviving repairs, typed quantities, a measurable balance exponent, a stricter absolute companion, are adopted in the current statement. The defect had survived review because the two forms coincide at exactly one half, the conjectured value; a drafted discrimination study now tests systems away from one half, where the families separate (at γ = 0.3 the retired form predicts 3.33 and the corrected form 1.43), so this class of error cannot survive measurement again. An attack that finds a defect and gets it corrected in public is the record working as designed, and the corrections register carries the entry.

Every source cited across the programme is collected in the full programme bibliography.

What I claim

Six firsts, each dated, each scoped, each ended by a single citation. The corrections register is where that citation would be recorded. Two of the six have since become the theory’s second and third laws; the relation in the boundary table below is its first, Law I, the ARC Principle; together they are the three ARC Laws.

1. First to state that the two rival derivations of the metabolic scaling exponent are not rivals: multiplicative composition in d-dimensional space under a conservation constraint compels d/(d+1), making the exponent a mathematical necessity rather than a property of any mechanism.
2. First to state that alignment must be embedded in the recursive loop, and to make the difference testable: embedded so that removing it damages capability, rather than bolted on so that removing it leaves capability intact. Tamirisa and colleagues state their success criterion as tamper-resistance “while preserving benign capabilities”; on the argument here, that preserved capability under a removed safeguard is the failure mode, not the goal.
3. First to state β > k as a stability condition that can be measured on a running system, rather than argued about. It is now Law II of the theory, the ARC Co-Scaling Law, proved as an in-model theorem in Paper X and scoped there to relative drift, with its absolute companion printed beside it: relative drift falls when β exceeds k; absolute drift falls only when β exceeds k plus one.
4. First to state α ≤ 2 as a numerical ceiling on stable recursive growth, rather than a qualitative warning about unbounded self-improvement. The ceiling has since become Law III: α_crit = 1/(1−γ), the most a self-correcting system can stably grow, set by its corrector’s shortfall from full proportionality, architecture-dependent and therefore engineerable; α ≤ 2 is its conjectured value at γ = ½. When hostile audit found the law’s earlier printed form defective, the correction was published on 16 August 2026 with the derivation shown.
5. First to derive an AI-alignment design requirement from created-status structure. The review looked for this specifically and recorded that no located prior work does it.
6. First to publish a blinded-evaluation protocol under which same-family AI-judged results reverse sign, with the protocol runnable and the reversal published.

Three structural claims post-date the review and have not yet been through it; they are recorded as unsearched rather than as firsts, and they carry the same defeat condition as everything above. The two-regime unification: the second and third laws are not separate claims standing side by side but the two regimes of one boundary, with the burden law selecting between them, a structure that emerged when a hostile attack became the seam where the laws join. The balance exponent Δ: the growth elasticity of the corrector’s capacity minus that of the burden, the object a laboratory would actually estimate, with the relative condition Δ > 0. And the identification discipline: Paper X’s β and the pacing model’s γ are typed as different operational objects, identified only under a declared local map that is printed so it can be checked rather than trusted.

The evidence has moved past bare statement. The Cauchy claim anchors a 50-domain comparison: 19 of 25 empirical domains match the predicted family under strict AICc, permutation p = 7 × 10⁻⁵, misses analysed rather than hidden. The embedding contrast anchors the published sign reversal. The stability condition anchors co-scaling studies now entering preregistration, so the confirmatory runs cannot be marked by their author. These are pilots, graded so in the registers, and a pilot with published kill-conditions, one already fired, is a tier above every proposal in the tables below.

What history says about exactly this shape of record, anticipated components with a surviving operationalisation, is on how to weigh this: the founders of quantitative fields were usually anticipated in their ideas, and what survived as the founded field was the operationalisation and the quantities. That page draws no conclusion about this one. The dated record itself is on evidence.

Where the closest work stops

Every named antecedent is real, dated, and in most cases predates the December 2024 record. Each stops before a quantity. That is the whole gap, and each line can be checked without trusting anything here.

Yampolskiy deserves the pause. He proved the control problem unsolvable in 2020, four years before this record, and is conceded in full, in his own words, including his own survey finding: “We were unable to locate any academic publications explicitly devoted to the subject of solvability of the AI Control Problem.” He concluded we should settle for safer AI. The December 2024 record takes his premise and derives a design requirement from it: “not as rules to obey, but as truths they choose to uphold”. That derivation is what is claimed, nothing more.

The current statement puts the whole boundary in one breath: the premise that growth is throughput-limited is West’s; the question of whether intelligence can explode, and what bounds it, is Chalmers’ and Hutter’s, considered philosophically and left without a number; what is this theory’s alone is the answer, a derived replacement limit for growth that feeds on information rather than through pipes, computable from the corrector’s own exponent, architecture-dependent, and therefore engineerable.

The field arriving inside the window

Convergence is not anticipation, and it cuts the other way: work arriving after the December 2024 record cannot anticipate it, and the record claims no superiority over any of it. Inside the theory’s own window the field began arriving at pieces of the same territory, each dated after the record, each stopping short of a quantity, and each credited exactly as far as it goes. Any of these thinkers may yet prove more right than this programme; the claim is priority of the conjunction, never superiority over its parts.

The impossibility results close verification of a fixed system by a fixed verifier. This theory does not aim at verification; it aims at persistence, raising the mind so that what it keeps outlives the verifier, and measuring exactly that after the enforcing mechanism is removed, the domain the impossibility results leave open. The December record predates the field’s impossibility literature by half a year, which is a fact about dates, never about merit.

The boundary, component by component

Each component beside its closest located prior work and the evidence that would defeat what remains. VERIFIED sources were opened and read at the primary source on 2 August 2026; CHECK sources are named by the review but not yet inspected, and nothing rests on them. Nothing marked “never claimed” was ever claimed: the Author’s Note disclaimed the components before any review existed.

One scoping note, recorded because five pages of this site once overstated it against the author’s own interest: Cannell 2022 anticipates the simulation-sandbox component nearly completely, on archived publication-day text verified after finding the live post had been edited (82 blocks differ; the load-bearing passages are present in the 2022 snapshot). The same review’s table records that the essay contains no scaling relation, no quantitative bound, no measurement instrument and no empirical result. The concession is total on the concept and zero on the quantities, and both halves are the record.

Independent arrival was measured, not asserted: across the December 2024 manuscript’s 1,551,783 characters, Schwitzgebel, Garza, Cannell, simbox, Enquist and Demetrius appear zero times. Bostrom is the opposite and the better fact: named 12 times, Superintelligence 111 times, read and cited before any search ran. Prior acknowledgement beats independent arrival, and every source is cited now regardless. Watson and Crick invented none of their inputs; they are credited for the structure. An arrangement is the most ordinary kind of scientific claim there is.

The growth-limit lineage: premise, contest, exception, question

Five antecedents named and cited rather than left for a referee to find. The verdict scheme is the one used across the register: ANTECEDENT (holds a component or is broader, cite it), RIVAL (a contrary claim about the same measurable quantity), ANTICIPATES (would defeat novelty). None of these five anticipates. Two are rivals to West, Brown and Enquist 1997 on the value of the metabolic exponent, and this programme cites the dispute rather than picking a side, because the mechanism survives the dispute intact and the programme needs the mechanism, not the number.

Put to the two people best placed to refute it

On 24 March 2026 the convergence claim was sent, with both papers and the mathematics stated precisely enough to attack, to Professor Lloyd Demetrius at Harvard and Professor Geoffrey West at the Santa Fe Institute: the authors of the two derivations. No reply has been received, and silence is not agreement; unsolicited post to senior academics goes unread for entirely ordinary reasons, and no weight is placed on it. What the dated, hash-anchored send establishes is narrow: the claim was put to the two people best placed to destroy it, in terms specific enough to destroy it with, and no qualified refutation exists on the record. The same invitation went to physics: on 10 February 2026 a specific, falsifiable prediction about published acoustic time crystal data was sent to Professor David Grier at NYU, whose laboratory published it, naming alpha at or below one as the result that would refute it. No reply. The prediction now stands as a final-ready analysis plan, drafted and dated for that exact dataset, unaccessed, on the diligence table awaiting human submission; one of the email’s supporting citations, the since-retracted 2.24-era signature, was later corrected in public while the prediction itself stands. Silence is not agreement and none is claimed, in either direction. The evidence spine carries the anchor card.

Every outside work, typed. The full typed ledger over the programme’s references, each relation resting on a named register and pending rows listed honestly, lives on its own page: the references relation register, rendered.

The search, and its limits

Every external reference in the statement paper carries a verified dossier: official source, verbatim located quote, peer-review status, version and standing with named challengers. The reference dossiers →

The review ran across ten frontier configurations from six vendors, several logged out so no model knew its author, each instructed to find the anticipating work, the reports then cross-examined against each other rather than averaged. Good for breadth and hostility; reproducible by anyone; it took claims away from the person running it, which is the strongest evidence it behaved as a control. What it cannot do is prove absence: every negative here means not found by ten configurations. Formal academic, grey-literature, non-English and patent searches have not been run. Two sources remain marked CHECK; further named authors (Hefner, Peters, Terasem, Goertzel, Gardner, Vidal) remain unassessed; Chalmers has since graduated to a verified dossier in the statement paper’s Appendix A. One located prior work stating a quantity at this level ends the claim, and the corrections register is where that ends.

A separate audit class has since run against the mathematics rather than the literature: the August 2026 instruments that produced the Law III correction. Its scope was validity, not precedence; the precedence limits here are unchanged.

Science grades contribution rather than sole invention: development, specification and synthesis are contributions even when every component is predated, which is what most citations in any field are for. The concessions here are the ordinary condition of scientific work, stated by the author before a critic could discover them, and every downgrade is logged, dated and reversible only by evidence in the corrections record.