2026-10-11 16:33 UTC

A. Dabrowski claims an AI-agent-produced paper and 238-module Lean 4 artifact prove the Hilbert-Smith conjecture, which would add a machine-checkable solution of a major mathematical problem if the formalization and underlying argument withstand expert review.

state: seedheat: mediumuncertainty: mediumnovelscott: mediumai-assisted-mathematics lean4-formalization hilbert-smith-conjectureA. DabrowskiMathlib

What is this?

The case centers on a claim by A. Dabrowski that an AI agent produced both a paper and a 238-module Lean 4 formalization proving the Hilbert–Smith conjecture — a long-standing problem about locally compact group actions on manifolds. Web search results do not surface this specific claim or author; they return an Apple research paper introducing 'Hilbert,' an agentic framework for informal-to-formal mathematical reasoning (arXiv:2509.22819, NeurIPS 2025 MATH-AI workshop), plus general discussion of Lean 4 auto-formalization benchmarks and a MathOverflow thread on trusting AI-generated Lean proofs. The Hilbert–Smith conjecture itself is not mentioned in the returned snippets. If the Dabrowski artifact exists, it is not indexed in the top results for this query.

Why it matters to Scott

The claim — an AI agent producing both a paper and a 238-module Lean 4 formalization of the Hilbert–Smith conjecture — sits squarely on multiple load-bearing Scott themes: AI-assisted mathematics, Lean 4 formalization workflows, agentic informal-to-formal pipelines, mathematical sovereignty/local-inference proof artifacts, and frontier machine-checkable solutions. However, the grounding search found no trace of this specific claim or author; the artifact appears unindexed and unverified. If it materializes, it would be a consequential milestone; as an uncorroborated claim it is a watch item, not yet a signal.
radar:concept.ai-assisted-mathematicsradar:concept.lean-formalizationradar:concept.formal-verificationradar:concept.automated-theorem-provingradar:concept.mathematical-discoveryradar:concept.frontier-modelsradar:concept.local-inferenceradar:concept.agent-harnessesradar:concept.agent-frameworksradar:concept.provenanceradar:lean-transformer-ai-proofsradar:proofatlas-collatz-formalizationradar:kbr-ai-formalized-math-proofradar:anthropic-fermat-lean-formalizationradar:qinz1yang-poincare-lean-formalizationradar:hopf-proof-codex-lean-formalization
queries asked of Scott's wikis
  • ai-assisted-mathematics lean4 formalization workflow
  • agentic-frameworks informal-to-formal math verification
  • mathematical-sovereignty local-inference proof-artifacts
  • lean4-mathlib integration ai-generated-proofs
  • frontier-math-problems machine-checkable-solutions

Measured heat

now 0 pts/hpeak 2 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 99h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion

How the heat travelled

10-07 13:00⭐ origin echo-reconstructedSelf-published proof of the Hilbert–Smith conjecture plus a Lean 4/Mathlib formalization, released the same day as the HN post. README: "**`
A. Dabrowski (GitHub: adbrw; HN: aldabrow) on github (echo) · attributed from hn.story.50003926
—
10-08 10:06first on hacker news · published · +21.1hIndependent Hilbert-Smith proof with Lean4 verification
aldabrow
—
10-08 10:06amplified on hacker news 👑hn.story.50003926
aldabrow
peak 2 · 1 comments · 101% of case engagement
10-08 10:34our radar first saw it · +21.6hdiscovery anchor: hn.story.50003926—
pace: p24 vs 1247 stories at the 96h mark (now 99h old) — ahead of aafp-commons-signed-agent-notebook (2.0x), behind agentgate-signed-agent-receipts (0.7x)

Evidence (2) — ⭐ canonical anchor

sourceobjectauthorscorecomments
🟧 hnIndependent Hilbert-Smith proof with Lean4 verification
Retrieved article excerpt

Open article · Retrieved 2026-10-08T10:42:37.963868+00:00

# Integral tail signatures and the Hilbert–Smith conjecture

**[`HSproof.pdf`](https://github.com/adbrw/HS_proof/blob/main/HSproof.pdf)**: *Integral tail signatures and the Hilbert–Smith conjecture*,
A. Dabrowski's AI agents, October 8, 2026, 17 pp.

> **Theorem 1.1** (p-adic exclusion). For every prime p, a continuous action of ℤ\_p on a connected
> finite-dimensional manifold has nontrivial kernel.
>
> **Corollary 1.2** (Hilbert–Smith). A second-countable locally compact Hausdorff group acting
> continuously and effectively on such a manifold is a Lie group with its given topology.

Manifolds are Hausdorff and second countable, and may have boundary. All actions are jointly
continuous. Corollary 1.2 follows from Theorem 1.1 by the standard reduction to p-adic exclusion,
using locally compact group structure and Newman's theorem [Pa19].

The obstruction is an integer signature tail. The proof has two parts:

- **Divisibility** (§§2–6, Theorem 4.3). Take cyclic groups G\_i = C\_{p^{a\_i}} with index-p
  subgroups P\_i, and a fixed compact free C\_p-space T. Tensoring with the positive permutation
  form τ\_i = ℚ[G\_i/P\_i] is locally p copies of the identity. Controlled Mayer–Vietoris over a
  finite cover makes p − τ ⊗ (−) nilpotent. Equivariant signature characters then show that
  every class in L\_4(A\_G(T)) has ordinary signatures eventually divisible by p.
- **Realization** (§§7–13, Proposition 13.1). An effective ℤ\_p-action gives a class with
  signature tail 1. Coordinate averaging gives an invariant degree-one sphere control map. Fine
  local chain models turn finite quotient representatives into a homotopy action. Free sheets
  cancel the detector displacement, and averaging gives a duality-compatible homotopy idempotent.
  Its normalized Poincaré corner forgets to the scalar manifold germ of Sⁿ × CP². Successive
  localization boundaries recover one copy of CP², up to sign.

The two parts together give 1 ∈ pℤ, a contradiction. The paper uses the following controlled
machinery from the literature: Karoubi localization [CP95], controlled Poincaré–Lefschetz duality
[Ran99], algebraic Poincaré pairs and unions [Ran80I, Ran80II], and idempotent splitting [BS01].
It works out the local carrier and normalization calculations this setting needs.

## Lean 4 formalization

The rest of the repository is a Lean 4 / Mathlib formalization of Corollary 1.2, together with
the scripts that check it.

`Challenge.lean` states the Hilbert–Smith conjecture as the theorem `hilbert_smith` (imports only
Mathlib, proof `sorry`). `Solution.lean` proves the same statement as `HSFormal.hilbertSmith n M G`.
`HSFormal/` (238 modules) is the proof: exactly the import closure of
`HSFormal/HilbertSmithNegK.lean`, where `HSFormal.hilbertSmith` is proved; `HSFormal.lean` is the
library root importing it. Axioms used: `propext`, `Classical.choice`, `Quot.sound`.

### Contents

| Path |  |
| --- | --- |
| `HSproof.pdf` | the paper |
| `Challenge.lean` | statement, `import Mathlib` only |
| `Solution.lean` | same statement, proof `HSFormal.hilbertSmith n M G` |
| `HSFormal.lean`, `HSFormal/` | the proof (library `HSFormal`); `HSFormal/Brouwer/LICENSE` is the MIT licence of the five `HSFormal/Brouwer/` files |
| `lakefile.toml`, `lake-manifest.json`, `lean-toolchain` | Lean `v4.35.0-rc3`; mathlib `1a547d8a48a8fa7877d2decb69d7294723bb0187`; TauCeti `c7af81f021f76fa11afcd61c894e5fc861ab6e78` |
| `comparator.json` | comparator configuration |
| `verify/build_tools.sh` | builds comparator, lean4export, landrun and SafeVerify at pinned commits |
| `verify/safeverify-v435.patch` | ports SafeVerify (Lean v4.27) to Lean v4.35.0-rc3 |
| `verify/run_comparator.sh`, `verify/run_safeverify.sh` | run the two checkers |

### Requirements

- Linux ≥ 5.19 with Landlock enabled (in a container, seccomp must allow the `landlock_*`
  syscalls), for landrun, comparator's sandbox. Comparator calls landrun with `--best-effort`:
  without Landlock ABI v2 (Linux < 5.19) the whole ruleset is silently dropped, and network
  restriction (≥ 6.7) and IPC scoping (Landlock ABI v6, ≥ 6.12) are dropped on older kernels.
  `verify/run_comparator.sh` first checks that a sandboxed write outside `.lake` is denied, and
  stops otherwise.
- [elan](https://github.com/leanprover/elan), git, curl (used by `lake exe cache get`), `which`
  (used by comparator), Go ≥ 1.24 (to build landrun).
- About 15 GB of free disk and 16 GB of RAM.
- Run as an unprivileged user (comparator's assumption 6).

### Running

From this directory:

```
lake exe cache get              # Mathlib and its dependencies, prebuilt
verify/build_tools.sh           # tools, into verify/_tools/
verify/run_comparator.sh        # expected last line: Your solution is okay!
verify/run_safeverify.sh        # expected last line: SafeVerify check passed.
```

Run comparator before anything compiles `Solution.lean` or `HSFormal` (comparator's assumption 2):
`verify/run_comparator.sh` deletes earlier build outputs of `Challenge` and `Solution`, and
comparator then compiles `Challenge`, then `Solution` with the `HSFormal` modules and the
`TauCeti` modules they import (those not already built), inside its landrun sandbox.
`COMPARATOR_SYSTEMD=1 verify/run_comparator.sh` runs comparator under
`systemd-run --property=RestrictAddressFamilies=~AF_UNIX --user --pty`, as its README prescribes
(needs a systemd user session).

To only compile the proof: `lake build` (builds `HSFormal`), then
`lake env lean --stdin <<< 'import HSFormal.HilbertSmithNegK #print axioms HSFormal.hilbertSmith'`.
aldabrow21
🟧 echo.github ⭐Self-published proof of the Hilbert–Smith conjecture plus a Lean 4/Mathlib formalization, released the same day as the HN post. README: "**`A. Dabrowski (GitHub: adbrw; HN: aldabrow)——

Interpretation history

Decision trace