Off by 0.8, With a Margin of Fourteen Anthropic on 17 November published a post titled "Claude solves two open problems in the physics of black holes," but the two accompanying preprints dated 7 November on Hexagon state that Claude was given 22 candidate open problems and solved two after roughly six weeks of agent work, with one earlier agent having failed in 61 parallel attempts on the bootstrap problem. The bootstrap preprint gives an exact answer of Λ = 12π²/(π² + 4 arcsinh² 1) ≈ 9.1266, tightening a slope previously known only to lie between 8.503 and 12, while the Hawking-Page preprint reports a coefficient b = −1.48 × 10⁻³ λ², making the transition first order. The Hawking-Page paper's Appendix C requires that "no number enters the result unless two independently written programs agree on it, or it has been checked against an exact result," a standard the author says the post's framing does not fully match. On 17 November Anthropic published a post titled “Claude solves two open problems in the physics of black holes.” Both problems come from quantum gravity research. The first asks whether a phase transition that Stephen Hawking and Don Page found in 1983 stays abrupt when spacetime has no definite shape. The second asks for the exact limit of a technique called the modular bootstrap. Behind the post are two preprints, both dated 7 November on Hexagon, a preprint server. Both list “Claude Anthropic ” as the author and carry the server’s labels “Primarily AI-generated text” and “Human understanding: some parts.” I’m not a physicist and can’t referee the physics. I can read how the papers say their numbers were checked and compare that with what the post says. The two don’t fully match. In one respect the papers are more modest than the post, and in another they are more honest, and the modest version is the one I’d trust. Two results, two kinds of claim The bootstrap paper is about a ceiling on how heavy the lightest black hole can be in a simplified three-dimensional universe. Physicists write the bound as Δ₁ ≤ c/Λ, where c roughly counts the universe’s microscopic building blocks. A larger Λ means a tighter ceiling. Simeon Hellerman’s 2009 paper published in 2011 started this line of work. By 2019 Thomas Hartman, Dalimil Mazáč and Leonardo Rastelli had proved a slope of 8.503. Nima Afkhami-Jeddi, Hartman and Amirhossein Tajdini pushed numerical searches to c = 1800, reached c/8.956 there, and extrapolated to about 9.08. The preprint states the open question plainly: “The best slope Λ was known only to lie between 8.503 and 12.” It then gives an exact answer, Λ = 12π²/ π² + 4 arcsinh² 1 ≈ 9.1266, and proves that this version of the method can’t do better. The other paper is a calculation, not a theorem. Ofer Aharony, Joseph Marsano, Shiraz Minwalla, Kyriakos Papadodimas and Mark Van Raamsdonk AMMPV showed in 2003 that whether the weak-coupling transition is abrupt depends on the sign of one coefficient, b, which first appears at three loops. They worked it out in 2005 for pure Yang-Mills theory but never for N = 4 super Yang-Mills, which the preprint says is “larger by more than a factor of ten.” The preprint’s answer is b = −1.48 × 10⁻³ λ². The coefficient is negative, so the transition is first order. The post also gives the denominator, which I always go looking for after writing about Claude’s enzyme campaign that worked once in eleven runs https://gautamparab.com/writing/one-campaign-in-eleven/ . An Anthropic researcher asked Claude for a list of open problems. It “came back hours later with 22 candidates.” Agents worked on them for about six weeks, and two were solved. On the bootstrap problem, an earlier Claude agent “had tried, and failed, to solve in 61 parallel attempts.” So it’s two problems out of twenty-two. That’s a good result for a side project. It’s also a different headline from “Claude solves two open problems,” because it says nothing about the other twenty. The rule: no number from a single program The Hawking-Page preprint describes its own quality control at length, more than most human-written papers do. Appendix C sets out three rules imposed from the start. The first: “no number enters the result unless two independently written programs agree on it, or it has been checked against an exact result.” The second: every program had to pass “a gate,” meaning a case with a known answer, run by the same code with nothing changed but the input. The third: “a disagreement is never averaged or argued away. It is pursued until one side is shown to be wrong.” The independence goes as far as method. One construction of the interaction vertices uses 9j symbols, the other uses explicit harmonic polynomials, and the paper says “they share no code.” Three engines were written for one piece of the contraction and two for the other. The paper also admits one exception. The coefficients of one series from the seventh order onward “were produced by one engine only,” and those unconfirmed terms contribute −0.34 to a quantity whose value is −1.50. I like this design. My problem is with the word “independently.” What software engineers found when they tested independence Software engineers tested this assumption forty years ago, and the answer was discouraging. In N-version programming you have several teams write the same program from one specification and let them vote, on the theory that independently written programs will fail independently. John Knight and Nancy Leveson checked that theory in 1986. They had 27 versions of a program written from one specification at two universities and ran each through one million tests. The programs were individually very reliable: six never failed, and 23 of the 27 passed more than 99.9 percent of tests. But in 1,255 tests more than one version failed on the same input. The test statistic was 100.51 against a 99-percent cutoff of 2.33, and the authors rejected the independence model. They also explained why. Their starting worry was that “people tend to make the same mistakes for example, incorrect treatment of boundary conditions even when they are working independently.” The data agreed. The faults differed, but their effects “tended to involve the same boundary conditions or unusual cases.” One set of collinear points broke eight programs. D. E. Eckhardt and colleagues ran a bigger version in 1991, with 20 versions of an aerospace program written and independently validated by 60 programmers at four universities. When the programs did fail, up to 8 of the 20 failed together. In those conditions a three-version voting system was only two to five times less likely to fail than a single version. Three of the shared faults came from “a common and apparently inadequate understanding” of the non-orthogonal coordinate systems in the specification. The paper’s conclusion was “a modest gain in reliability.” The baseline matters for the black-hole paper because of who wrote its programs. Knight and Leveson’s programmers had different teachers and different habits, and they still made correlated mistakes on the hard cases. The preprint’s own appendix says the work “was divided among many instances of the same AI system, several at a time, each with a narrow brief and no memory of the others except through files.” Separate briefs and separate files make the programs independent as code. They are not independent in the sense that counts here. Copies of one model share one training history, so whatever that history gets wrong about a hard step, say how to treat a contact term or a sign convention in a counterterm, they are likely to get it wrong the same way. Two engines agreeing on a number is evidence, but much weaker evidence than two research groups agreeing. The paper’s gates help with exactly this. A known answer doesn’t care who wrote the code. The gates include the su 2 sector, matrix models with N = 2 and 3, AMMPV’s two-loop results, and the two-loop Hagedorn temperature known from integrability. A shared blind spot would show up there, provided the blind spot affects the gate cases. The strongest test in the paper is the one place where the three-loop machinery meets a number computed by humans: AMMPV’s 2005 pure Yang-Mills result. The row that didn’t agree Section 8 reruns the method on pure Yang-Mills and compares it with the 2005 paper diagram by diagram. AMMPV’s fourteen three-loop diagrams fall into seven classes. The logarithmic coefficients “agree exactly, diagram by diagram.” Five classes agree within rounding, and a sixth agrees to within one unit in the last digit of one entry. The seventh class, two Coulomb interactions plus a sextic vertex, gives −2.095 in the new calculation against −3.388 in AMMPV’s entries. That one class accounts for the gap between the new pure-glue value, b = −4.3 × 10⁻⁴ λ², and the 2005 value of −5.7 × 10⁻⁴ λ². The paper’s response is the passage I’d most like other labs to copy: “We have not resolved this.” It sets out what each side has. On its side, three codes for one part of the class, and two engines with independently written ordering remainders that agree to 36.9553 5 against 36.957 6 . On AMMPV’s side, a recomputation of two of the three diagrams with a different regulator that “did not change.” It points out that the 2005 paper prints the expression for every diagram and points to a Mathematica file, so someone can push the comparison further. And it ends with “One of the two calculations has an error in this class.” By the preprint’s own account, AMMPV wrote in 2005 that they had no independent way to verify their number. Twenty years on there are two numbers that disagree, and the new paper says so in plain words. The paper argues that the conclusion survives anyway, and the arithmetic is simple enough to check. In the units it uses for the three-loop term, the computed value is q₄ = −2.234, and it “would have to be +11.8 to change the sign” of b. That is a margin of about fourteen. If AMMPV turn out to be right about the seventh class, the shift is 0.8 in those units, and it goes toward a more negative b, so further from the boundary. The single-engine terms contribute 0.34. I accept the margin argument as far as it reaches. It covers the errors the paper found. It can’t cover an error that every engine shared and no gate tested, which is exactly the kind Knight and Leveson found in their 1,255 coincident failures. What would settle it is a human group, or at least a different system, redoing the seventh class from AMMPV’s Mathematica file. The paper says this in its own way: the disagreement “should be resolved in the future.” Proof, check, and a line about Lean The bootstrap paper handles certainty differently, and I think better. Every statement carries a status tag. “Proved” means proved in the paper from textbook facts. “Proved, computer-assisted” means proved modulo finitely many inequalities “certified in interval arithmetic: rigorous floating-point arithmetic with outward-rounded error bounds.” Then there are “conjecture,” “heuristic,” “numerical” and “orientation.” The slope 9.1266 is a theorem. The coefficient of the next correction, tested by numerics up to c = 2048, is labeled a conjecture. When the paper mentions a related result proved elsewhere, it says “we have read the statements and the structure of that argument but have not independently verified it; nothing in this paper depends on it logically.” The post says something the paper doesn’t: “In September, the results were formally verified in Lean, a computerized proof checker.” I searched both preprint PDFs and their Hexagon pages for the word “Lean” and found it nowhere. The paper names interval arithmetic and a table of certified inequalities as its machine checking. A Lean formalization may well exist and simply not be cited, but I couldn’t find it, and the paper that’s meant to be the record doesn’t claim it. If the formal proof is real, publishing the files would be a bigger contribution than the post. Anyone can check a Lean proof on a laptop, and nobody can check a sentence in a post. What I’d take from it These are two different kinds of claims, and the post presents them as one. The bootstrap result is a theorem, its status tags tell a reader how much rests on computer arithmetic, and a specialist can check it line by line. The Hawking-Page result is a large calculation whose main check is agreement among programs written by copies of the same model. It passes several gates with known answers, and it has one open disagreement with the only earlier human calculation of the same kind. The paper’s case that the disagreement can’t flip the sign holds up, as far as I can tell. The quote I keep coming back to is “a disagreement is never averaged or argued away.” That’s the right rule. The N-version experiments add a second rule: agreement between copies of the same system is cheap, and a check means more the less the checker has in common with the thing being checked. On that measure, the 2005 paper by five human physicists is the most valuable check in either preprint, and the one place it was used, it found a problem. I’d want to see that problem resolved before calling the first result settled. References 1. Anthropic 2026 . Claude solves two open problems in the physics of black holes https://www.anthropic.com/research/claude-black-hole-physics . Published 17 November 2026. 2. Claude Anthropic 2026 . The Hawking-Page transition at weak coupling https://hexagonmath.org/2611.00175v1 . Hexagon 2611.00175v1, 7 November 2026. 3. Claude Anthropic 2026 . The spinless modular bootstrap at large central charge https://hexagonmath.org/2611.00174v1 . Hexagon 2611.00174v1, 7 November 2026. 4. Aharony, O., Marsano, J., Minwalla, S., Papadodimas, K. and Van Raamsdonk, M. 2003 . The Hagedorn/Deconfinement Phase Transition in Weakly Coupled Large N Gauge Theories https://arxiv.org/abs/hep-th/0310285 . arXiv hep-th/0310285, 31 October 2003; Adv. Theor. Math. Phys. 8 2004 603. 5. Aharony, O., Marsano, J., Minwalla, S., Papadodimas, K. and Van Raamsdonk, M. 2005 . First order deconfinement transition in large N Yang-Mills theory on a small S³ https://doi.org/10.1103/PhysRevD.71.125018 . Physical Review D, 71, 125018, June 2005. 6. Hellerman, S. 2011 . A universal inequality for CFT and quantum gravity