The open problem
Evidence
The dated record, each entry carrying the check that would break it
One law may govern how everything grows. Physical things are held below unbounded growth by the geometry of the space that feeds them. Software is the first thing built outside that geometry. A self-improving mind stays stable only if its correction grows as fast as its capability. Everything below is that claim in testable parts, with dates, kill-conditions and checks.
What was seen
Six firsts, each dated, each ended by one citation.
- 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).
- Alignment must be embedded in the recursive loop, testably: removing it damages capability, or it was never embedded.
- β > k as a stability condition measurable on a running system.
- α ≤ 2 as the numerical ceiling that condition implies at the limit of how well any corrector inside the system can scale: if correction could scale without limit, there would be no ceiling at all.
- An AI-alignment design requirement derived from created-status structure.
- A blinded-evaluation protocol under which same-family AI-judged results reverse sign, published with the reversal.
The strongest pilot so far: fifty domains in total, of which twenty-five carry empirical curve fits. The composition operator was classified before fitting. 19 of those 25 empirical domains match the predicted family under strict AICc, permutation p = 7 × 10−5. Misses analysed in Paper VII. Exploratory, and graded so; independent replication decides the rest.
When
8 December 2024. A five-attachment manuscript package, 221,236 words, self-emailed and preserved as the ARC-sealed recipient-form record, with OpenTimestamps receipts anchored 6 August 2026. ARC mathematics passed against live keys on 2 August 2026. Inside: the embedded-correction thesis, the control-failure passage, and the conversion heuristic U = I × R (the law form U = I × Rα adds the measured exponent) stated about ninety-five times with all three terms defined. The Universe as the totality of existence. Intelligence as the organising principle spanning evolution and artificial systems. Recursion as the mechanism that perpetuates creation. Framed, already, as the complement to the mass-energy law. SHA-256 begins f0d1f38f. Verify it yourself.
30 April 2025. The complete book manuscript, self-emailed and preserved as the ARC-sealed recipient-form record, carrying the Eden Protocol by name, thirty-seven concepts, and the first appearance of the squared form on the dated record.
2 January 2026. Infinite Architects published: 468 pages, ISBN 978-1-80605-620-0, checkable in any catalogue.
24 March 2026. The convergence argument sent to Professor Lloyd Demetrius at Harvard and Professor Geoffrey West at the Santa Fe Institute, the authors of the two derivations it connects, with the mathematics stated precisely enough to attack. A dated, hash-anchored record; no qualified refutation exists on it. See the fuller account on related work.
What would kill it
The kill-conditions are published with live statuses, each ending its claim the day it fires. One has fired: the early sequential-scaling exponent of 2.24 failed replication, the corrected 0.49 (interval [-1.3, 2.9]) replaced it in public. The newest, which the programme raised against itself, is that same 0.49: it sits almost exactly on the 0.5 independent-sample averaging predicts with no framework at all, a second null distinct from the pinned-ratio null below.
The zero-parameter null. The architecture-blind alternative is not free to fit: under the specified exchangeability null, the expected cross-class to same-class ratio is exactly 1.00, with no adjustable parameter. The framework predicts a ratio materially above 1.00; the null’s expectation cannot move, and the registered analysis carries the confidence interval and equivalence margin that decide between them. A prediction whose competitor is pinned is the cheapest kind to kill, which is why the confirmatory measurement is drafted for registration before anyone runs it. The trigger figure on the consequences page and the falsification register both carry it, and the retraction is on the record. The rest are open. Any of them ends its claim the day it fires. The live dashboard.
Check it
- Hash the published files in your browser and resolve the timestamps through independent resolvers.
- Read the corrections log: what somebody does when the evidence goes against them.
- 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 broadcast record: two BBC broadcasts preserved and dated, the full chronology printed cold beside the same-day correspondence, and every recording’s hash on the page.
- The machine-readable registers, generated from one source and parity-gated against these pages.
- How to weigh work of this shape, which draws no conclusion.
- The related work register: what each named antecedent stops before, with the differential for each.
The evidence here is what exists. the dated prediction register is what was said in advance about what the evidence would show, which is the harder half to produce and the easier half to check.
One record here was written as the work happened rather than about it afterwards. The laboratory notebook ran live from 10 March 2026 to 206 pages, and it keeps the failures in: the false positives blinding exposed, and a list of the author’s own mistakes made at the moment he found them.