Archived edition

Updates on artificial intelligence for mathematical research, proof discovery and formal verification, with links to papers, code and tools.

Research updatesSelect an entry to read more

2 entries · 7 linked sources Partial coverage
Axiom announces a Lean formalization of the Stacks Project Axiom Published: 4 sources

Axiom announced a 16-million-line Lean library intended to cover the approximately 7,500-page Stacks Project and link declarations to permanent Stacks tags. The public explorer is potentially useful for algebraic geometers as a navigation and reuse layer: it explicitly separates StacksCore, described as human-reviewed and Mathlib-standard code, from StacksFrontier, described as AI-agent-written code awaiting that review. This is a genuine new release, not a repetition of the earlier CFSG announcement.

Research relevance:

A large formal corpus for algebraic geometry could make dependencies, reusable lemmas, and machine-checkable statements substantially more accessible for Lean-based research and autoformalization.

Claim status:
Announcement and claimed completed formalization. Axiom states that Lean's kernel checks completed proofs; the public project site confirms the stated Core/Frontier distinction. No independent audit or proof review was found, and I did not run Lean.
Limitations:
The announcement is first-party. The project site does not by itself establish fidelity of every Lean statement to the corresponding natural-language Stacks result. The pinned GitHub repository could not be consulted through the available web cache, so code contents, buildability, completeness, and the stated scale were not independently checked.

Sources & further reading

Harmonic reports Lean-backed verification of Ibex multiplication and division instructions against Sail Satnam Singh and Harmonic Published: 3 sources

Harmonic published an engineering report claiming that Aristotle proved nine properties for the RV32M single-cycle Ibex multiplier/divider configuration: eight result-equivalence assertions for MUL, MULH, MULHSU, MULHU, DIV, DIVU, REM, and REMU, plus an instruction-completion condition intended to avoid vacuity. The comparison uses a Sail RISC-V model, includes division by zero and signed-overflow cases, and the report says the SystemVerilog assertions are compiled to Lean 4 against Harmonic's circuit model. For researchers using AI agents, it is a concrete workflow example: a natural-language request, generated harness, specification comparison, and reported proof obligations.

Research relevance:

This is a practical autoformalization and proof-engineering case study at the boundary of theorem proving and hardware verification. Its scope is narrower than full processor verification, but the explicit non-vacuity condition and edge cases are useful details for evaluating such claims.

Claim status:
First-party engineering announcement claiming machine-checked formal proofs. Harmonic's public documentation describes Lean-backed verification, but the complete run report requires sign-in and no public proof artifact or independent replication was located. This is not an independently checked proof claim.
Limitations:
The stated result concerns the Ibex mult/div module under a specified configuration, not the whole processor or all RISC-V extensions. It depends on the supplied RTL, the selected Sail/SVA formal specification, and Harmonic's formal circuit embedding. I did not inspect the sign-in-only report, execute the tool, or independently verify the Lean artifacts; Harmonic's claim that an existing Ibex model-checking setup omits MUL correctness was not independently audited.

Sources & further reading

Coverage and limitations
  • Time window checked: 2026-11-03T21:15:02+01:00 to 2026-11-05T09:15:01+01:00. I read Axiom's November 4 announcement and Stacks formalization site, Harmonic's November 3 engineering report and Aristotle documentation, and the Ibex documentation. I also searched for related primary sources and recent posts from Tao's blog.
  • Direct X collection: API status was ok for __alpoge__, scottnarmstrong, wtgowers, sebastienbubeck, thomasfbloom, littmath, alexkontorovich, f_charton, leonard41111588, boazbaraktcs, leanprover, projectnumina, harmonicmath, axiommathai, and epochairesearch. The collection was bounded, excluded replies and reposts, may contain cached posts, and API account resolution does not independently verify identity. Only the supplied API posts were treated as directly accessed X material; no complete X timeline was inferred.
  • Direct Bluesky collection: xenaproject.bsky.social status was ok but returned no posts; it was bounded to at most 300 posts and excluded replies and reposts. This does not establish inactivity. Mathstodon collection for tao@mathstodon.xyz was unavailable because of an HTTP error; replies and reposts would also have been excluded.
  • The Axiom Stacks repository link exposed by the project site, https://github.com/AxiomMath/Stacks/tree/f8e70c5dd04bfabcb423d5e325461f527c85a5d6, could not be read through the web cache. I did not execute Lean, inspect the sign-in-only Aristotle report, or independently audit either formalization. No suitable additional pure-mathematics result or proof announcement was found in the window; promotional statements and an unsupported post about an unreleased model were excluded.

Requested coverage window: 2026-11-03T21:15:02.113976+01:00 — 2026-11-05T09:15:01.435653+01:00.

Public web research; coverage is not exhaustive. A source link is not a certification of a claim.

Archives 79
Older editions (49)