All groups are good: the four-dimensional disc-embedding conjecture

All groups are good: the four-dimensional disc-embedding conjecture

OpenAI's claim

Manuscript 1 (family 305, 2026-09-24): A compact smooth four-manifold is claimed to contain disc maps with framed algebraic duals whose boundary circles admit no disjoint locally flat fillings, even without replacement framings or relative classes prescribed. Manuscript 2 (family 305, 2026-09-24): The unrestricted four-dimensional disk-embedding assertion is claimed false: algebraically suitable immersed disks lack disjoint locally flat replacements preserving boundaries and induced framings, even when relative homotopy classes may change.

Mathematical statement of the claimed result

Mathematical claim 1.

OpenAI claims a smooth oriented four-manifold $M$ that is compact and connected, with boundary nonempty, proper generically immersed disks $f_i:(D^2,S^1)\to(M,\partial M)$ with disjoint embedded boundary circles, and framed sphere maps $g_i:S^2\to M$, for some $k\ge1$, with \[ \lambda(f_i,g_j)=\delta_{ij}\,1,\qquad\lambda(g_i,g_j)=0,\qquad\widetilde\mu(g_i)=0, \] but with no pairwise disjoint proper locally flat embedded disks filling these circles. Here $\lambda$ is the equivariant intersection pairing in $\mathbb Z[\pi_1M]$ and $\widetilde\mu$ is the reduced equivariant self-intersection invariant. Each sphere can individually be smoothly embedded, although the spheres may intersect one another. The nonexistence conclusion imposes neither output framings nor dual spheres. Further corollaries assert that the rank-two free group $F_2$, and every group containing it, is not good in the Freedman--Quinn disk-embedding sense, and that unrestricted four-dimensional topological surgery and five-dimensional $s$-cobordism cannot both hold.

Mathematical claim 2.

For some smooth oriented four-manifold $M$ that is compact, connected and has boundary and integer $k\ge1$, OpenAI claims proper generically immersed disks $f_i$ with disjoint embedded boundaries and framed generically immersed spheres $g_i$ satisfying \[ \lambda(f_i,g_j)=\delta_{ij}\,1,\qquad\lambda(g_i,g_j)=0,\qquad\widetilde\mu(g_i)=0. \] The pairing $\lambda$ takes values in $\mathbb Z[\pi_1M]$; the reduced self-intersection invariant takes values in its additive quotient by $h-h^{-1}$ and integral multiples of $1$. Nevertheless no pairwise disjoint locally flat embedded disks have the same boundary maps and the normal boundary framings induced by the $f_i$. The preserved framings are part of this nonexistence statement.

Outline of the argument

Manuscript 1 (family 305, 2026-09-24): A cut construction retains boundary holonomy data from hypothetical fillings and converts them into formal gradient identities. Finite algebraic specialization passes these identities to positive characteristic, where truncated complexes expose an incompatible tensor pairing. Because the obstruction concerns the boundaries themselves, it also rules out the weaker unframed filling demand and implies that the rank-two free group is not good.

Manuscript 2 (family 305, 2026-09-24): Independently conjugated Brunnian relations are installed before cutting the hypothetical embedded images. Based flat SL_2 connections give Chern--Simons potentials, and algebraization preserves their full marked critical schemes, including nilpotents. After positive-characteristic truncation, transverse graph classes pair nontrivially, whereas triple relations and independently lifted parameters force every corresponding pairing to vanish.

Formal verification

Family 305 has no Lean scope page in this snapshot; no formalization of these manuscript results is listed.

This site has not executed the Lean files or checked their dependencies. These descriptions report the scope of the linked declarations, rather than a new verification of the proofs.

Status caveat

OpenAI attributes these results to an internal model. Its release notice cautions that results lacking full formalization can contain mistakes. This site has not independently assessed the proofs; a solved label records the scope of the reported claim.

Sources

• OpenAI, A boundary-only obstruction to four-dimensional disk embedding, 2026-09-24; citation key OAI:A-boundary-only-obstruction-to-four-dimensional-disk-embedding-September-24-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main; Statement source: cor:nongood; Statement source: cor:surgery-conjunction.
• OpenAI, A marked tensor obstruction to four-dimensional disk embedding, 2026-09-24; citation key OAI:A-marked-tensor-obstruction-to-four-dimensional-disk-embedding-September-24-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• OpenAI catalogue entry.

All groups are good: the four-dimensional disc-embedding conjecture

Definitions and mathematical statement

A capped grope is a surface tower built by attaching surfaces along symplectic basis curves and capping its tips by disks, with the standard four-dimensional thickening and framing. Height $1.5$ refers to the source's half-stage attachment convention. A double-point loop of an immersed disk follows paths on its two sheets through a self-intersection and is based using a path in the thickening. The selected definition calls $\Gamma$ good if every based height-$1.5$ disk-like capped grope and every homomorphism $\phi:\pi_1(G^c)\to\Gamma$ admit an immersed disk with the prescribed framed boundary whose double-point loops all lie in $\ker\phi$. The conjecture is that every group is good. The source's equivalence to goodness of $F_2$ is retained in the quotation; its disk-embedding consequence concerns locally flat topological disks, not arbitrary smooth embeddings.

Formulation and concept references: [S1].

Short English statement

Can every prescribed capped-grope configuration be replaced by an immersed disk whose double-point loops become trivial in the target group, for every group?

Research attempt

Scope and literature input.

GPT-6 Astra Ultra prepared this unresolved attempt on 27 September 2026 using the 2026 primary problem list [S1], group-word calculations and checks of the disk quantifiers. The source's Section 4.5 defines goodness through framed immersed disks in capped-grope thickenings and records closure properties and known subexponential-growth cases. Its definition is stronger than algebraic cancellation of a signed intersection sum. The argument below proves elementary reductions, reconstructs the reduction to two generators using the specified known closure input, and tests why finite quotients or abelianization do not finish it. No smooth disk-embedding theorem is claimed.

What is required for one fixed grope.

Fix a based finite height-$1.5$ capped grope $G^c$, its attaching region, and a homomorphism $\phi:\pi_1(G^c)\to\Gamma$. The attaching circle is null-homotopic in the thickening; the standard framed immersed-disk existence in this setup is part of the source's discussion. Let $g_1,\ldots,g_r$ be the based double-point loops of one such immersed disk. The target is to choose a disk for which \[ \phi(g_i)=1\quad(1\le i\le r). \] A different choice of basing paths conjugates the relevant group elements, and exchanging sheets inverts them. Membership in $\ker\phi$ survives both operations because that kernel is a normal subgroup. Thus the condition is geometrically meaningful despite those choices. It is not a request that a preferred written word have zero exponent sum.

If $\phi$ is the trivial homomorphism, every double-point loop already passes the test. This settles that instance even when $\Gamma$ itself is complicated. More generally, it is the actual image of the particular homomorphism, rather than every element of the ambient target group, that controls the application of a known goodness theorem.

Two direct inheritance lemmas.

If $\Gamma$ is good and $H\le\Gamma$, then $H$ is good. Given $\phi:\pi_1(G^c)\to H$, compose it with the injective inclusion $H\hookrightarrow\Gamma$ and apply goodness. The resulting disk's loop images are trivial in $\Gamma$, and injectivity makes them trivial already in $H$. The disk, attaching region and framing are unchanged in this argument.

There is a useful instance-wise variant. If the homomorphism factors as \[ \pi_1(G^c)\xrightarrow{\psi}\Lambda \xrightarrow{\eta}\Gamma \] with $\Lambda$ good, then goodness for $\psi$ supplies the required disk for $\phi=\eta\psi$. Triviality before applying $\eta$ is enough; $\eta$ need not be injective. This is a factorization criterion, not a proof that every map to a quotient admits such a lift. In general, relations in the domain can obstruct lifting a homomorphism through a surjective group map.

Since the capped grope is a finite complex with a standard thickening, its fundamental group is finitely generated. Therefore the image of $\phi$ is a finitely generated subgroup of $\Gamma$. If every finitely generated subgroup of $\Gamma$ is good, factor through that image and apply the preceding criterion. This proves that it suffices to settle all finitely generated groups, without assuming compact four-manifolds realize arbitrary infinitely presented groups.

Why the free group on two generators is the universal test.

The nontrivial established closure input from [S1] is that quotients of good groups are good. I do not derive it merely from the factorization lemma: an arbitrary map into a quotient has not been lifted by that lemma. With this known closure result in hand, the reduction to $F_2=\langle a,b\rangle$ is elementary.

For each $r$, the elements \[ a^jba^{-j}\qquad(1\le j\le r) \] freely generate a subgroup of rank $r$. To check independence, write a nonempty reduced word in these generators, grouping consecutive powers of the same generator. Between powers of distinct generators, expansion leaves a nonzero power of $a$ separating nonzero powers of $b$. The free reduction cannot remove all those $b$ blocks. Thus no nonempty reduced generator word becomes the identity in $F_2$. If $F_2$ is good, the subgroup lemma makes every $F_r$ good. Every finitely generated group is a quotient of an $F_r$, so quotient closure makes it good. The preceding finite-image argument then handles an arbitrary target group. The reverse implication is immediate because $F_2$ is itself a group.

This reconstruction clarifies the content of the selected universal statement: it is not made easier by restricting to finite presentations. It also identifies exactly which part of the reduction rests on the known four-dimensional goodness theory, namely quotient closure. Free-group word calculations alone have not produced the required disks.

A growth collision calculation, and its geometric limitation.

Suppose a geometric construction produced $2^h$ distinct histories at stage $h$, each labeled by a word of length at most $ch$ in one fixed finite generating set of a target group. Write $v(R)$ for the number of group elements in the word ball of radius $R$. If the group has subexponential growth, then for sufficiently large $h$, \[ v(ch)<2^h. \] Indeed choose $\epsilon<(\log2)/c$ when $c>0$ and use $v(R)\le e^{\epsilon R}$ for all sufficiently large $R$. The pigeonhole principle gives two histories with the same label; the quotient of their label words is trivial. The case $c=0$ is immediate because the ball has one element.

In $F_2$, the positive words of length $h$ in $a,b$ are all reduced and distinct, giving $2^h$ collision-free histories. Its full symmetric word ball has size \[ 1+4\sum_{j=0}^{h-1}3^j=2\cdot3^h-1. \] There is consequently no corresponding collision forced by this count. This explains a real algebraic difference between the growth regimes. It does not prove that absence of a forced collision obstructs every possible disk construction, or that a collision alone supplies one. To use the estimate geometrically one must produce the histories with the claimed length control and convert the collision into framed disk moves without introducing uncontrolled new loops. Those requirements are not proved here. The known subexponential result in [S1] uses geometric work beyond this elementary counting observation.

Why finite quotients cannot be interchanged with one disk.

Suppose one applies known goodness of finite groups to every finite quotient $q:\Gamma\to Q$. The conclusion has the logical form \[ \text{for every }q\text{ there exists a disk }D_q \text{ whose loops are killed by }q\phi. \] The desired conclusion asks for a single disk with loops killed by $\phi$. Residual finiteness distinguishes each nontrivial fixed loop, but here the disk and its loop list may change with $q$. It does not allow these two quantifiers to be exchanged.

There is an explicit word-level stress test in $F_2$. Given any finite collection of finite quotients $q_1,\ldots,q_N$, let $m$ be the least common multiple of the orders of $q_j(a)$. Then $a^m\ne1$ in $F_2$ but every $q_j(a^m)=1$. Thus every finite collection of quotient tests can miss a nontrivial word that is allowed to depend on the tests. This is not asserted to be a list realized by a particular capped-grope disk. It refutes the bare inference that arbitrarily many finite tests, with changing candidate data, give simultaneous triviality in the group. One would need an additional compactness, bounded-complexity or coherent geometric selection theorem, none of which is supplied by residual finiteness alone.

Intersection sums and framing do not encode every loop condition.

Replacing the target by its abelianization loses another essential piece of information. The commutator $aba^{-1}b^{-1}$ is nontrivial in $F_2$ but has zero image in its abelianization $\mathbb Z^2$. Making all double-point loops homologically trivial would therefore not make them trivial in the specified group. Similarly a signed intersection expression can contain the formal cancellation $g-g=0$ in $\mathbb Z[\Gamma]$ with $g\ne1$. Vanishing of that sum does not say that either individual loop has trivial image. These are algebraic countertests of proposed criteria, not constructed counterexamples to goodness.

Geometrically, pairing intersections suggests Whitney disks, but in a four-manifold a two-dimensional Whitney disk and a two-dimensional surface have complementary dimensions. General position allows isolated new intersection points; it does not make the disks disjoint from the surfaces automatically. The boundary framing is an additional condition which a move must preserve or repair. A higher-dimensional Whitney-trick argument that discards these intersections and framings would miss the specific difficulty of the four-dimensional problem. The locally flat topological embedding consequences in the source are likewise not statements about arbitrary smooth embedded disks.

Verification and outcome.

The subgroup and finite-image reductions check the exact kernel condition for each double-point loop. Free-word reductions and the word-ball formula were checked finitely as illustrations of their all-length proofs. The finite-quotient example explicitly retains a nontrivial word and does not claim geometric realization. The first remaining gap is a framed immersed-disk construction for every required capped-grope map into $F_2$, with all double-point labels genuinely trivial. Neither counting, residual finiteness nor algebraic intersection cancellation establishes it. The all-groups-good conjecture remains UNRESOLVED.

Sources

• S1 Baykur, Kirby and Ruberman (eds.), K3 – A New Problem List in Low-Dimensional Topology, preliminary2026author version; Section4.5 scribed by Powell and Ray. https://bpb-us-e2.wpmucdn.com/websites.umass.edu/dist/b/22144/files/2026/04/K3-problem-list-watermarked.pdf. Formulation source.