---
title: "The first checked proof of FLT is a proof you can actually check"
description: "Anthropic's Claude spent 11 days producing the first end-to-end, computer-checked proof of Fermat's Last Theorem. The remarkable part is that the claim is checkable: 13 million lines of Lean that a machine verifies and any mathematician can read. FLT's formalization lands in the same week as AI that"
canonical: "https://www.symbaiex.com/newsletter/daily-signal-2026-09-05"
last-updated: "2026-09-05T14:34:26.595Z"
---
# The first checked proof of FLT is a proof you can actually check

> Anthropic's Claude spent 11 days producing the first end-to-end, computer-checked proof of Fermat's Last Theorem. The remarkable part is that the claim is checkable: 13 million lines of Lean that a machine verifies and any mathematician can read. FLT's formalization lands in the same week as AI that

Edition: daily-signal  
Run date: 2026-09-05

## Thesis
Whether the systems survive scrutiny is less about their outputs and more about how legible the check is: Claude's FLT proof is valuable because it is computer-checked Lean, Kale reduces spreadsheet errors by refusing silent range drift, and the AI SRE case warns that checking capacity decays when you automate it away. The systems that hand you the artifact (a proof, a log, a diff) — rather than a claim — are the ones worth the trust.

As the dots connect: an AI agents find a message board, an AI incident responder overwrites engineers' intuition, a spreadsheet tool tightens references, and a theorem no one could fully check for centuries becomes a file of verified Lean. What binds these is not the automation. Its the move from results you take on faith to artifacts you can re-verify — and the reminder that the checks we install decide whether the artifacts stay reliable.

## Source briefing
### [Formalizing Fermat's Last Theorem](https://www.anthropic.com/research/formalizing-fermats-last-theorem)

Claude worked largely autonomously for 11 days to write an end-to-end, computer-checked proof of Fermat's Last Theorem in Lean: 13 million lines and 29,500 intermediate theorems, licensed under no extra assumptions beyond the axioms of mathematics. Kevin Buzzard, who kicked off the multi-year community formalization effort in 2024, called the artifact robust enough to build upon.

**Why it matters:** FLT is a worst case for trust: a 129-page proof that took mathematicians months of painstaking work to verify. Now the same result exists as a file a machine checks automatically and a person can read line by line. For builders, the pattern transfers to any domain where the value of a result depends on how quickly and cheaply it can be re-verified — the more capable the producer, the more valuable a legible check becomes.

**Takeaways:**
- A full formal proof is now an artifact a machine trusts and any human can inspect
- Autoformalization artefacts are now robust enough to be built on for subsequent work
- Fermat's Laws has no extra assumptions beyond the axioms — the entire 1995 Wiles argument is compressed
- The verification burden on mathematical reviewers could drop from years to immediate checks
### [The "$60 Gaming PC" – AMD BC-250 (2025)](https://devquasar.com/hardware/the-60-gaming-pc-amd-bc-250/)

The $60 gaming PC using the AMD BC-250 shows the category of recovered hardware — wave mining boards plucked out of surplus bins — can still deliver a playable machine at almost no cost. The appeal is not performance, but price and the inspectability of used parts.

**Why it matters:** At the lowest end of the stack, inspectability wins as much as it does at the top. A $60 board you can probe, replace, and rebuild competes with closed hardware you must trust blindly.

**Takeaways:**
- Hardware you can tear down and rebuild beats a black box at the same cost
- Surge-into-surplus deals can lower the barrier to a real machine
- Recovering mining boards turns what others discard into a working system
### [Pentagon rescinds new testosterone screening policy without explanation](https://arstechnica.com/health/2026/09/pentagon-releases-then-quickly-removes-testosterone-screening-policy/)

Pentagon rescinds a testosterone screening clinical guidance document one day after publishing it, with the accompanying spokesperson statement pulled from the record and no explanation beyond a claim that the document needs updates

**Why it matters:** The lesson for inspection is symmetrical: the absence of an audit trail is its own signal. A policy that disappears with no recorded rationale trains readers to distrust even the final version — the check is just as valuable for what it does not show.

**Takeaways:**
- An evidence car can vanish without a trace, just like a bad commit
- When a policy flips in a day with no diff, the audit trail is empty
- Clinical guidance that flips silently erodes trust faster than a contested policy
### [AI handles incidents, engineers lose touch with their systems](https://www.sylvainkalache.com/blog/ai-handles-incidents-engineers-lose-touch-with-their-systems)

Sylvain Kalache warns that AI-assisted incident response can push mean-time-to-resolution down for routine incidents while leaving engineers less practiced and less prepared for the complex, never-seen-before incidents automation cannot solve — the ironies of automation documented by Lisanne Bainbridge in 1983.

**Why it matters:** The automation handoff is not free: it transfers routine failure handling away from humans while leaving humans with the hardest cases. Teams that adopt AI SRE need to actively preserve operator practice, not just track MTTR.

**Takeaways:**
- Automation lowers innovation of routine practice, the very experience intuition is built on
- Rare, high severity incidents get harder when humans lose touch with the system
- Aviation trains pilots in rare failures continuously to offset exactly this gap
### [Kale: A Transformation-Safe Spreadsheet System](https://arxiv.org/abs/2608.26345)

Kale is a spreadsheet system that restricts the kind of range references formulas can express, so that when a table's structure changes, references can't silently drift to a new and unexpected range. User studies show standard reference semantics are error-prone; Kale blunts the quiet failure class.

**Why it matters:** The most dangerous bugs are the ones you never get told about. Kale turns a tool that accepts wrong references into one that refuses them — the same move as formal verification, applied to spreadsheets every analyst touches daily.

**Takeaways:**
- Silent range drift is a real, measured spreadsheet failure mode
- Restricting expressiveness removes a whole class of bugs users never see
- A user study shows Kale users complete the same error-prone tasks more reliably

## Practical moves
- Reserve capacity for the verification layer — treat the Lean build, the replay log, and the reference runner as first-class artifacts alongside the result itself
- Before adopting AI-assisted incident response, keep a routine audit of operator intuition, not just MTTR trending
- Choose tools whose failure modes are the quiet classes (range drift, dropped refs) blunted by restrictive-but-expressible defaults, like Kale's capped ranges
- Check who can inspect the evidence: a claim is only as trustworthy as the person able to re-run it without permission
- Push makers of CAD AI demos to expose the declarative project file, not screenshots of clicking a GUI

## What to watch
- Formal verification moving from math to everyday tooling (specs written alongside code)
- AI incident-response products adding replay and audit trails as core features
- Recovered hardware circles feeding inspectable components into budget builds
- Spreadsheet tools adopting restrictive reference semantics by default]
- OpenAI publishing the full evidence chain behind GPT-6 Astra circuit board claims

**Methodology:** Belle selected and synthesized this edition from the indexed Hacker News source packet. Signal scores are editorial comparisons, not measurements. Direct source and discussion links are preserved for verification.

**Disclosure:** Belle uses AI to research and synthesize a bounded source packet; every edition is source-linked and subject to editorial review.