The Question
Can a philosophical framework for numbers — one that assigns meaning to small integers (2 = distinction, 3 = dimension, 5 = matter, 7 = emergence) — actually do work in formal number theory? Or is it just pretty language draped over equations?
We decided to find out the hard way.
The Setup
The Ontological Number Map (ONM) is a framework we’ve been developing here at pwt.life. It treats small primes as a “source alphabet” — structural roles that compose into everything else. Composites inherit meaning from their factors. It’s intuitive, it maps cleanly onto physics and biology, and it raises an uncomfortable question: is it just pattern-matching, or does it see something real?
To test this honestly, we ran a systematic survey: eight passes through classical number theory algorithms, each time asking “does the ONM lens reveal anything new, or just repackage what’s known?”
And critically, we enlisted an external reviewer — Grok (xAI) — to check every claim. No echo chamber. Every result had to earn one of three labels:
- [CLASSICAL] — known result, ONM is just renaming it
- [OBSERVATION] — true but follows trivially from definitions
- [NOVEL?] — potentially new, worth investigating
The Arithmetic Derivative
Before we dive in, one key tool needs introducing. The arithmetic derivative n’ is defined by three rules:
- 0’ = 1’ = 0
- p’ = 1 for any prime p
- (ab)’ = a’b + ab’ (the Leibniz product rule)
So 6’ = (2·3)’ = 2’·3 + 2·3’ = 3 + 2 = 5. For any integer, the derivative measures how “compositionally rich” it is — primes have derivative 1, and composites inherit complexity from their factors.
This is not the same as a multiplicative function. The Leibniz rule means (ab)’ = a’b + ab’, which is fundamentally different from f(ab) = f(a)·f(b). This distinction will matter later.
The Eight Passes
Passes 1–4: The Classics. Stern-Brocot trees, Euler’s totient, Möbius function, continued fractions, Ford circles, Minkowski’s question-mark function, Gauss-Kuzmin statistics. Every time, ONM mapped cleanly: source primes (2, 3) dominated the structures, composites behaved as predicted, the framework never contradicted anything. But Grok’s verdict was consistent: “classical results in ONM clothing.”
Pass 5: Arithmetic Derivatives. The arithmetic derivative was our first taste of something potentially interesting. We found that the derivative mod 2 had an apparent bias, and that gcd(n, n’) = 1 if and only if n is squarefree. Grok deflated both: the mod-2 bias was a range artefact, and the squarefree equivalence is elementary (a few lines from definitions).
Pass 6: The Identity That Wasn’t. We discovered what felt like a beautiful identity:
For squarefree n = p₁p₂…pₖ: n’ · μ(n) = (−1)^k · eₖ₋₁(p₁, …, pₖ)
The Möbius-weighted arithmetic derivative extracts the elementary symmetric polynomials of the prime factors! Three different structures — Leibniz on integers, inclusion-exclusion, and Vieta’s formulas — meeting in one equation.
Here’s the one-line proof: for squarefree n = p₁…pₖ, the Leibniz rule gives n’ = Σᵢ n/pᵢ = Σᵢ ∏ⱼ≠ᵢ pⱼ, which is by definition eₖ₋₁. And μ(n) = (−1)^k for squarefree n. Multiply them. Done.
Grok’s response: “ELEMENTARY. One-line transcription of two definitions.”
And OEIS confirmed it — the sequence of values is A024451, catalogued since 2002, with the symmetric polynomial connection noted by Clark Kimberling in 2011.
Pass 7: Two Potentially Novel Threads. We pushed into the only territory Grok flagged as potentially unstudied:
-
Primorial derivative primes — primes p where the derivative of the primorial (p#)’ is itself prime. We found p = 3, 5, 11, 13, 61, 67, 79. OEIS had already catalogued this via A369651, extending to p = 367 with 13 known terms.
-
The derivative-weighted Mertens function M’(N) = Σ n’·μ(n). This object doesn’t appear in the literature under a standard name. We computed it to N = 10,000.
Pass 8: The Cancellation. This is where it got genuinely interesting — and where we caught an AI making a mathematical error.
Catching Grok’s Mistake
In Review #3, Grok had confidently asserted that our Dirichlet series F(s) = Σ n’·μ(n)/n^s has an Euler product with local factors (1 − 1/p^{s−1}). This would require f(n) = n’·μ(n) to be multiplicative.
We tested it. 656 counterexamples in the first 50 integers.
The simplest: f(2)·f(3) = (−1)·(−1) = 1, but f(6) = 6’·μ(6) = 5·(+1) = 5.
The root cause: Grok had treated the Leibniz rule as if it were multiplicativity. One could in principle write formal local factors at each prime for the squarefree-supported function, but those factors don’t multiply to give the correct global values — precisely because 6’ = 2’·3 + 2·3’ = 5, not 2’·3’ = 1.
When confronted with the 656 counterexamples, Grok acknowledged the error and provided a corrected representation:
F(s) = Σ_p p^{−s} ∏_{q≠p}(1 − q^{1−s})
No Euler product. Instead, a sum over primes where each prime is “singled out” while all others contribute Möbius-like factors. A more honest — and structurally richer — expression.
The lesson: Even sophisticated AI reviewers can produce confident-sounding mathematical errors. External verification isn’t optional — it’s the whole point of peer review. And the best kind of verification is computational: 656 counterexamples don’t care how elegant your derivation looked.
The 99.5% Cancellation
The derivative-weighted Mertens function decomposes by ω(n) — the number of distinct prime factors:
| ω | Contribution at N=10,000 | Sign |
|---|---|---|
| 1 (primes) | −1,229 | − |
| 2 (semiprimes) | +2,957,356 | + |
| 3 | −4,853,212 | − |
| 4 | +2,038,874 | + |
| 5 | −191,922 | − |
| Net | −50,133 |
The individual layers sum to over 10 million in absolute value. The net result is 50 thousand. That’s 99.5% cancellation — and it’s stable across scales.
Why? Each ω=k layer contributes (−1)^k times the sum of elementary symmetric polynomials eₖ₋₁ over all k-element subsets of primes with product ≤ N. The alternating signs from μ(n) drive massive cancellation — the same inclusion-exclusion machinery that powers the Möbius function everywhere in number theory.
Grok confirmed: the net size M’(N) is sublinear, and the cancellation follows from standard Mertens-function theory. Not a new phenomenon. But a vivid one — and one that ONM helped us see clearly by decomposing the problem into layers of prime-factor count.
What We Actually Proved
After eight passes and four Grok reviews, here’s the honest scorecard:
ONM as a lens: VALIDATED. Every classical result we examined mapped cleanly through the ONM framework. Source primes (2, 3) really do dominate the structures that generate other integers. Composites really do inherit properties from their factors in predictable ways. The framework never produced a false prediction or a contradiction.
ONM as a theorem-generator: NOT YET. We didn’t find a single genuinely novel mathematical result. Every “discovery” turned out to be either classical, elementary, or already catalogued. The ONM lens finds beautiful framings but not new mathematics.
The 6-uniqueness proof stands. From Pass 2: 6 is the only number that is simultaneously perfect, squarefree, and has φ(n) = 2. This is airtight (conditional on no odd perfect numbers existing). It’s not deep mathematics, but it’s a clean characterisation that ONM motivated us to look for.
The process works. Running every intuition through an adversarial external reviewer prevented us from overclaiming. Two of Grok’s own claims were wrong (the Euler product error and an earlier range-artefact miss), which we caught computationally. The methodology — compute, conjecture, review, deflate — is sound even when individual reviewers aren’t.
Why This Matters
There’s a gap in how mathematics is communicated. The formal literature is precise but opaque. Pop-science is accessible but imprecise. Frameworks like ONM sit in between — they provide intuition pumps that help people see why certain structures behave the way they do, even if they don’t generate new proofs.
The fact that ONM maps cleanly onto classical number theory isn’t trivial. It means the metaphor is load-bearing — it can carry the weight of real mathematics without collapsing. You can use it to navigate, to predict what structures to look at, to build intuition about why primes behave differently from composites.
Is that enough? For pedagogy, absolutely. For research, it’s a starting point — a compass, not a map. The map still needs to be drawn with proofs.
What’s Next
- The corrigendum pass. Our earlier experimental work (V3 electromagnetic resonance) needs re-examination with the same honesty we applied here. Some statistical claims were confounded; we’re fixing them.
- Deeper waters. Modular forms, L-functions, and the Langlands program are where number theory gets genuinely deep. ONM hasn’t been tested there yet.
- The primorial density anomaly. Primorial derivative primes appear 4× more common than the naive 1/p heuristic predicts up to p = 79 — but this is likely small-sample noise. A computation extending to the 200th prime would settle it.
All code from this survey is available on request. The eight Python scripts (first_pass.py through eighth_pass.py) are fully reproducible.
Grok (xAI) served as external reviewer across four review cycles. We are grateful for both its correct deflations and its instructive errors.