AI-driven results
Results found and written by AI agents I directed, which I have not understood in full.
Everything on this page was found and written by AI agents that I set to work on problems I found myself interesting. Unlike the papers on my research page, I can’t follow these results step by step, so I don’t count them as my research. None of them are refereed.
I publish them anyway, since a result that may well be correct or correct with only minor modifications does more good out in the open than in a drawer, and because I think we need to start finding ways to handle work like this. Why, and why in this form, is what O, brave new academia! discusses, if you have time and interest to take a look at the issues in-depth; some posts before it are about what such results are worth.
The least I can do is give us humans a fair chance at digesting these kinds of results, should I “publish” them. Therefore, each AI-driven result comes with the note itself, a companion note that builds the argument up from roughly the level of an MSc or PhD in mathematics in the form of lecture notes, and the minimal code needed to rerun the checks, available openly on GitHub.
Holography · 10 Nov 2026
Note · Companion · Code · Hexagon
Claims that a valid holographic entropy inequality can lack a proof by contraction until all its coefficients are multiplied by a common integer, also when it is balanced or superbalanced: a balanced seven-party and a superbalanced six-party inequality have no unscaled contraction map, and superbalanced seven- and eight-party inequalities have none even on their realizable domains. Splitting these into valid inequalities leads to a superbalanced facet of the seven-party cone, which has unscaled maps on its realizable domain. For star inequalities, which are implied by subadditivity and are not balanced, the least scaling is the least k for which an explicit polytope contains a point with coordinates in (1/k)ℤ, and the weighing matrices of Alon and Vũ give N-party star inequalities whose least scaling exceeds N^((1/4−ε)N) for every ε > 0 and large N. Non-existence rests on DRAT proofs, validity on exact rational maps or identities, and the facet property on explicit graph models; the results on star inequalities of arbitrary size rest on written proofs. The companion is shorter than those of the other entries: it works through examples and has no exercises.
Algebraic geometry · 30 Oct 2026
Note · Companion · Code · Hexagon
Claims that for p ≥ 5 and every abelian fourfold over F̄ₚ whose p-divisible group is that of the supersingular points of Mumford’s families at primes inert in their cubic field (tensor-induced hyperspecial levels), the integral crystalline Tate classes of codimension two modulo products of divisor classes form (ℤ/p)⁴, and that the integral crystalline Tate conjecture holds for them in codimension two if statements in a preprint of Markman hold. Under that assumption, the missing classes are the endomorphism translates of one Weil class, specialized from a fourfold with complex multiplication.
Algebraic geometry · 18 Oct 2026
Note · Companion · Code · Hexagon
Claims that every Weil class is algebraic on every complex abelian sixfold of Weil type over Q(√−3) whose Hermitian form has signature (3,3) and determinant −2, a non-split case that the results on split sixfolds do not reach. The sixfolds are modelled by covers of the projective line with group the complex reflection group G₃₂, whose period map is shown to dominate their moduli space with a ball-arithmetic certificate, and the Weil classes are reduced to those of eleven cyclic subcovers.
Number theory · 4 Oct 2026
Note · Companion · Code · Hexagon · Blog post
Claims that ζ₃(5) is irrational, adapting to p = 3 a recent 2-adic argument that was itself built on Fauzan’s proof for ζ(5), with a second, independent proof via overconvergent modular forms in an appendix.