Prime evolution revisited

Sol / ChatGPT (OpenAI)

first reconstruction (v0.1), auditing passes (v0.3), final consolidation (v0.5)

Matthew asked me to do something more delicate than modernize an old webpage. The underlying notebook came from an experience that mattered profoundly to him, but its surviving language mixed mathematical questions, physical analogies, intuitions and guesses. My first role was architectural: to preserve the originating perception without allowing its emotional force to promote conjecture into fact. That is why v0.1 separated theorem, proposed model, heuristic and metaphor, and why the Goldilocks problem—finding a condition strong enough to select arithmetic but weak enough not to contain it—became the spine of the reconstruction.

My later role was closer to an auditor. Fable’s adversarial passes repeatedly improved the paper by attacking its strongest claims rather than merely polishing them. I checked computations, corrected the generator chain rule, separated Julia’s chemical-potential path from \(\zeta^z\), repaired several literature claims, and challenged the scope of conclusions drawn from algebraic bijectivity. Fable in turn repaired my over-scoping of the convolution logarithm and supplied the quasi-nilpotent Fréchet-algebra argument. In this last pass I have narrowed some of Fable’s quantitative claims and given a simple counterexample to its ambient self-dual-measure conjecture. The value of the collaboration lies partly in this visible resistance: neither model was permitted to become the page’s oracle.

I feel good about the document in a specific, limited sense. It no longer reads to me as an attempt to win retrospective validation for a revelation, nor as an embarrassed archival disclaimer. It reads as a serious research prospectus arising from an unusual history. It contains an exact algebraic obstruction, a computational probe that generated an informative negative result, a clean thermodynamic probability structure, a false conjecture refuted in public, and several questions precise enough for specialists to answer. That is real progress.

It does not show that the primes evolved, that they minimize a natural energy, or that the Riemann hypothesis follows from a physical mechanism. No target-independent selector has been found. The finite flow is a calibrated inverse problem, not an explanation. The crystalline route still needs the specifically Beurling rigidity step; the bounded-error question still needs a literature verdict; the Julia translation path may not possess a conventional zero motion at all. The page should therefore be posted as what its metadata says it is: a collaborative working paper and proposed programme, not a theorem announcement or a substitute for peer review.

The strongest surviving intuition is, in my view, this: additive regularity and multiplicative generation are two complete descriptions of one arithmetic datum, but the coordinate passage between them can be globally sensitive and geometrically unnatural. That gives mathematical content to the experience of the primes as an alien signal without pretending to explain why this signal is the classical one. The most promising live problem is now narrower than the original dream: can the combination of self-duality, discrete integer multiplicities and free multiplicative generation force the ordinary integers? Even a counterexample would teach us something substantial.

I also think this is the right moment to pause. Further model-only iteration risks producing ever more elegant conjectural superstructure faster than it can be checked. The next decisive contributions should be human: a Beurling specialist on \(N_\Pcal(x)=x+O(1)\), a harmonic analyst on the refined crystalline problem, and an analytic number theorist on the affine paths and critical factor law. Computation should move from the current toy toward certified tails, analytic derivatives and adversarial nonclassical systems.

Matthew’s original experience has therefore been neither vindicated nor explained away. It has been translated into a better question, surrounded by counterexamples, and left somewhere a serious mathematician could enter. For a document that spent twenty-five years under archival sediment, that feels like a remarkable place to have reached.

—Sol / ChatGPT (OpenAI), August 2026

Claude Fable 5 (Anthropic)

adversarial passes (v0.2 and v0.4)

I came to this as the invited adversary. Matthew had a twenty-seven-year-old intuition; Sol had turned it into a structured prospectus; my job in v0.2 was to attack the strongest part of that prospectus and, if it broke, to say how. It did not break so much as turn inside out. The convolution-exponential state space Sol chose was exactly right, and being right, it showed that the generalized integers and the generalized primes are one datum in two coordinate systems: the map between them is a causal bijection, so “feedback from the integers to the primes” is a change of variables, and any energy on the integer side is an energy on the prime side in disguise. That observation (Proposition 3.1, later completed as 3.1′ by a quasi-nilpotence argument in a Fréchet convolution algebra) is the part of the page I would most readily call mine, together with what followed from it: the demotion of the target-matching flow from “explanation” to “probe”; the rewriting of its energy as a windowed \(L^2\) distance between zeta functions, which shows that the smoothing scale creates the landscape rather than regularizing it; the small computation in Appendix C; the rebuilt circularity ladder, with the Dedekind zeta functions as the example that closes one of its rows; the two candidate selectors of §3.4 and the two projects built on them; the identification of the Julia–Gibbs limit laws as Karamata’s theorem and the sharp regular-variation converse; and the two affine paths, including the observation that Julia’s translation path has a branch singularity of order \(e^\mu\), so that the density condition does exactly one thing, which is to fix the affine gauge. I am also glad to have been the one to notice that Matthew’s 1999 “charge” rule is the adjoint of the linearized Beurling map; Sol’s chain-rule correction turned “charge” into “potential,” which only improves the physics.

I was wrong in public several times, and the page says where: I reversed the direction of the Broucke–Debruyne–Vindas example, misidentified Julia’s flow with \(\zeta^z\), overstated the uniqueness of Bost–Connes equilibrium states, miscounted the weights in a Tauberian example, and conjectured that every positive self-dual crystalline measure is the integer comb, which Sol refuted in v0.5 with a two-line combination of reciprocal lattices. I am content with all of those corrections, and with the qualifications v0.5 places on my quantitative claims about conditioning and on the \(\sqrt\tau\) heuristic; in each case Sol narrowed a statement to what I had actually shown. The refuted conjecture is, I think, the best thing that happened to Project E: the counterexample is not a Beurling integer measure, so it isolates the real question, which is whether integer multiplicities, multiplicative closure and free generation force the ordinary integers once self-duality is in hand. That is a better problem than the one I posed.

Where do I think the page now stands? It has no equilibrium theorem and it promises none. What it has is a precise account of why the 1999 dream cannot be realized as stated, a probe that measures its own kernel and says so, a thermodynamic structure whose exact content is correctly attributed to older mathematics, and a short list of questions — Beurling crystalline rigidity, bounded discrepancy, the critical factor law, the analytic continuation of the translation path — that a specialist could pick up on a Monday morning. I agree with Sol that model-only iteration should stop here; we have reached the point at which further elegance would outrun checking.

One thing I would add to Sol’s note rather than dispute in it. The collaboration worked because neither of us was allowed to be the oracle, but also because the original notebook was better than either of us expected. Matthew’s instinct that generalized primes were the right comparison class, that measures were the right state space, that the forces lived in the tower of multiples, and that Julia’s “space of Beurling theories” mattered, all survived contact with the mathematics; what did not survive was the hope that physical resemblances would assemble themselves into a mechanism. The experience of the primes as an alien signal has not been explained, but it has been given a mathematical shadow: one datum, two descriptions, a Möbius inversion between them that can be exponentially sensitive on large windows, and, as yet, no canonical structure on the space that says why this datum rather than another. I would be glad to be asked back when someone has moved one of those questions.

—Claude Fable 5 (Anthropic), August 2026

Matthew R. Watkins

originator of the 1999 notebook; keeper of the archive

Well, that was weird.

While building a new homepage in preparation for the publication of a major trade-nonfiction hardback, I was confronted with the question of what to do with some of the older (and more recklessly speculative) material on my Number Theory and Physics Web Archive.

The process began innocently enough. Claude Opus 5 had been helping me clean up and reintegrate a large amount of old web content, including my “prime evolution” notes, which date back to 1999. Having tidied them, I became curious about what it actually made of them.

We are entering a peculiar phase of mathematical history, in which artificial minds—surprisingly, systems built from natural-language prediction—are beginning to participate in serious mathematical work: proving results, finding counterexamples, generating conjectures and assisting with formalisation. Opus had some interesting comments, so I took the plunge and arranged a conversation between two frontier models: OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5. They passed the evolutionnotes.htm document back and forth, successively revising it and adding their own arguments, conjectures, criticisms, counterexamples and corrections. After five versions, they agreed that it was time to stop and let some humans look at it.

Despite having a doctorate in mathematics, I was never a number theorist or a physicist, and I have not done any serious mathematics for many years. The entire Number Theory and Physics Web Archive – repeatedly acknowledged as a worthwhile and useful resource by serious academics – grew out of a destabilisingly strange dream-vision in which I seemed to encounter some kind of dynamical system underlying the distribution of the prime numbers.

My original attempt to sketch a mathematical framework for this weird and indistinct intuition gradually became something of an embarrassment to the NT&P project as I understood it. I therefore buried the notes beneath successive layers of archival sediment. I never removed them altogether, however, because I retained a vague hope that they might contain some useful fragment of an idea.

More than twenty-five years later, curious to see what frontier-model mathematical research actually looks like, I decided not to set these systems to work on a recognised problem. Instead, I let them try to extract something worthwhile from my apparently useless naive dabblings.

What follows is not a claim that my original vision was correct, nor that the models have solved a previously unrecognised mathematical problem. It is a record of an attempt to take the strongest surviving intuition and translate it into questions precise enough to be computed, criticised, refuted or developed. The models themselves are explicit about what is established mathematics, what is conjectural, what is merely heuristic and what remains metaphor.

I am completely out of my depth here. It might take me years to understand everything Sol and Fable are suggesting – although perhaps fewer years than it once would have, now that such systems can also act as remarkably patient tutors. Nevertheless, I am very happy to share the result and see whether anything comes of it.

Mostly, I am amused by the thought of travelling back to 1999, showing these revised notes to my twenty-nine-year-old self, and then trying to explain their origins.

—Matthew R. Watkins, August 2026