Claude Raised the Proven Share of Zeta Zeros on the Critical Line to 67.2% — and the Method Matters More Than the Number

On August 10, Anthropic published a research note whose author is a large language model. The result: at least 67.2% of the nontrivial zeros of the Riemann zeta function lie on the critical line Re(s) = 1/2. Unconditionally. The previous proven record was 41.6% (5/12), a number that had survived four decades of Levinson-method refinements.

Claude did not prove the Riemann Hypothesis. The paper says so itself, twice. What it did — and this is the part worth slowing down on — is turn a conditional 67% into an unconditional 67% with a method that is genuinely new: no zero-density estimates, no zero-free region, no mollifier. The arithmetic input is exactly Montgomery’s prime-side evaluation of the pair-correlation second moment at bandwidth ≤ 1, which has been unconditional since 1973. Everything that used to require RH to read has been replaced by linear algebra on a finite matrix.

Why this number has weight

The Riemann Hypothesis says every nontrivial zero has real part exactly 1/2. Unproved since 1859; a Clay Millennium Problem with a million-dollar bounty. The partial results are lower bounds on the fraction of zeros on the line, and the lineage is worth stating in full:

Hardy (1914): infinitely many zeros on the line.
Selberg (1942): a positive proportion.
Levinson (1974): at least 1/3, via the mollifier method.
Conrey (1989): 2/5, refining Levinson.
Bui–Conrey–Young, Feng, Pratt–Robles–Zaharescu–Zeindler: 5/12, the standing record until this week.

Every number in that list comes from the same family of techniques — Levinson’s method and its descendants. The new 2/3 does not. It comes from Montgomery’s pair-correlation circle of ideas, which was previously only usable conditional on RH. That is a structural break, not an increment: it is the first time the “RH-conditional pair-correlation consequences” have been separated from RH itself at this strength.

The number 2/3 has its own history. Montgomery showed in 1973 that, under RH, at least 2/3 of the zeros are simple (later sharpened to 67.9% by Goldston and others). The 2/3 here is the same constant, now unconditional — but for zeros on the line, and counted as distinct zeros, not as “simple” in Montgomery’s sense. That distinction matters: N₀* counts distinct zeros on the critical line, so multiplicities are not counted against you. The paper also proves at least 5/6 of all zeros are distinct, which follows from the same machinery.

Two more structural points. First, the theorem is about dyadic ranges [T, 2T] with an o(1) error; the liminf over T→∞ is at least 2/3, and the paper reports the finite-T ratio sits below its limit because convergence is slow (their Remark 5.9). Second, the method is not stuck at 2/3: with the optimal Montgomery–Taylor window the constant becomes 0.67250…, which is exactly the RH-conditional Montgomery–Taylor constant. The 2/3 is the clean statement; the method can do slightly better for free.

A different engine

Here is where the result stops being a record and becomes a technique. Weil’s explicit formula turns a positivity statement about test functions into a claim about zeros: the associated quadratic functional is positive semidefinite if and only if RH is true. That equivalence is both the power and the trap — every naive “prove RH by positivity” route dies on it, and Claude’s coordinator agent reports that in the earlier thousand-agent session, every proposed route’s first substantive step turned out to be Weil positivity or an equivalent in disguise.

The trick is to stop asking about the whole infinite-dimensional form. Claude restricts the pairing to a finite-dimensional space of test functions — a “Gabor system” of modulated windows, d ≈ λN(T,2T) of them — and analyzes the resulting finite real symmetric matrix. Now:

zeros on the critical line contribute positive semidefinite rank-one blocks;
off-line zeros pair as {ρ, 1−ρ̄} under the functional equation and contribute indefinite (1,1) blocks — a Kreĭn-space signature, not a Hilbert-space one;
Sylvester’s law of inertia bounds the signature of the whole form;
a rank–trace inequality, proved via von Neumann’s trace inequality, connects the signature to computable trace moments;
the traces themselves are evaluated unconditionally from primes up to X = (T/2π)^λ using Montgomery–Vaughan.

The zero side and the prime side are matched by a Poisson-summation identity that gives the sampling kernel exactly, with no aliasing. Numerically the two sides agree to 10⁻⁸ at small heights. The step that matters most is conceptual: RH was classically needed to read the zero side as a positive sum over real ordinates. Here the indefiniteness of off-line pairs is absorbed by linear algebra — you never need to know how far off the line a zero is, only that its (1,1) block is bounded in rank. The negative index is handled by the inertia theorem instead of being zeroed out by an unproved hypothesis.

This is also why Bombieri’s 2000 paper, “Remarks on Weil’s quadratic functional in the theory of prime numbers, I,” is the right ancestor to cite: he studied the finite truncations of the Weil form and showed that if RH fails with only finitely many bad zeros, the number of negative eigenvalues of a large enough truncation equals half the number of bad zeros. The signature of the finite compression encodes how RH fails. Claude’s contribution is showing that you can extract the on-line count from that signature without ever resolving the question of which off-line zeros exist.

What it doesn’t say, and where the ceiling is

The paper is unusually explicit about its limits. No certificate of this kind can exceed 0.68185 — the first/second moment information at bandwidth λ ≤ 1 has a hard ceiling, and λ ≤ 1 is essential to the argument (beyond it you need information the unconditional prime side doesn’t give). Reaching 0.70/0.80/0.90 would require pair-correlation support out to roughly 1.04/1.26/1.70, which means genuinely new arithmetic input, not more clever linear algebra.

So the honest read: this is not a step toward proving RH, and the authors don’t claim it is. It is a step toward the older and arguably more tractable project — the “consequences of RH made unconditional” program, in the spirit of what Bombieri and others have done for primes in short intervals. And it demonstrates that the pair-correlation machinery, once freed from RH, is stronger than the mollifier machinery at this particular task.

The 54-hour session behind it

The appendix is the more interesting read, and it should be read as a workflow paper, not a color story. A coordinator agent ran ~60 isolated subagents over a day and a half: 2,400 shell commands, hundreds of Python scripts, 31 million output tokens. The coordinator did almost no mathematics itself; its role was direction, triage, and verification.

Two subagents produced the key ideas. E2 proved the negative-index route was empty — the honest finite negative index is identically zero — and then inverted the idea, lower-bounding the positive index from prime-side traces to claim at least half the zeros on the line. E2-pairs spent 91 minutes thinking, ran numerical experiments that contradicted its own brief, and wrote the rank–trace lemma that lifted 1/2 to 2/3. Four minutes after writing it, an infrastructure fault killed the run mid-sentence. The coordinator read the dead agent’s directory, recognized the result as “the most important thing to happen tonight,” checked the five-line proof line by line, and resumed the same agent with a checklist.

A few things stand out about how this worked. The coordinator’s own two creative forecasts were wrong — it pointed the mechanism the wrong way at the one-half step and gave a wrong recovery lever at the two-thirds step — while its referee architecture held. The subagents were deliberately isolated: they couldn’t see the conversation or each other, which kept their errors independent and made adversarial review meaningful. And the human’s contribution was a handful of one-line prompts: “Resume your work,” “Push it to 2/3,” “Keep going.” Claude initially underestimated itself; the persistence came from outside.

The verification protocol is the real finding

This is where the result matters beyond number theory. The trust problem for AI mathematics is not “can a model find a proof” — it’s “why should anyone believe it.” The pipeline here is a template for answering that:

Blind hostile referee agents, each assigned a disjoint gap and an attack plan, forbidden to read one another. They found one real error — a false premise about the mass matrix — and proposed the repair that ended up in the paper. An adversarial loop caught a hallucination; that is the system working.
A blind re-derivation agent that proved the result without reading the proof, and control cases where RH is known to fail. Both are cheap forms of reproducibility.
A literature agent that downloaded 54 arXiv papers to check novelty — the coordinator explicitly refused to trust its own memory.
A Lean 4 / Mathlib formalization of Theorems A–E, sorry-free, depending only on the three standard axioms (propext, Classical.choice, Quot.sound). That includes Weil’s explicit formula, Riemann–von Mangoldt counting, Stirling estimates, Chebyshev–Mertens, and Montgomery–Vaughan — the analytic scaffolding formalized too, not just the headline theorem.
Human experts: Anthropic’s in-house mathematicians Levent Alpöge and Ralph Furman validated it; external experts Brian Conrey and Dan Goldston examined it on short notice.

The coordinator’s own hedges are the most instructive part. “I’m not telling you half the zeros are on the line. I’m telling you an agent produced an argument with that conclusion.” “The result needs a human expert.” The final draft’s verdict: “The next reader should be a person.”

That is the right division of labor, and it’s worth naming it: generation and execution parallelize; verification is layered and adversarial; formalization is the ground truth; judgment stays with people. The error the referees caught is the reason the pipeline exists — a model generating a plausible proof is precisely the case where an uncritical reader gets burned.

The honest caveats

This is a preprint-level research artifact, released the same week it was written. The expert note itself flags unexpanded error terms (O(T^δ log T) and O(T^{1/2−2δ}) tails sketched rather than fully detailed). The Lean formalization is a static research artifact, not a maintained library. And the session-level behaviors — a coordinator whose forecasts were wrong twice, an agent that needed an infrastructure crash to be noticed — are not yet a reproducible process; they’re a data point about what this workflow can produce when it works.

The math has been checked harder than most preprints: a machine-checked Lean proof, an adversarial referee loop, and two independent expert reads. Final judgment still belongs to the community, which is how it should be. But the shape of the thing — an AI-generated theorem, formally verified, expert-reviewed, with its method honestly mapped to its ancestors in Bombieri and Montgomery — is new. The number 67.2% is the least surprising part of the announcement.