Claude built a proof checker, then found its own flaws
After building a minimal proof assistant, the writer asked Claude how they knew the prover was actually sound. Claude checked, found five different ways it could prove false theorems, and fixed them.
Done withClaudeSpotted· by HN user kragen· Oct 21, 2026
Result Also proposed a fuzzer that turned out useless
“It checked and immediately found five different ways it could "prove" false theorems with it (and fixed them).”
Spotted: found in the wild. Details are as reported by the source above and have not been independently verified.
What they did
The writer had built a minimal proof assistant with Claude (Fable 5.1). When Claude thought the project was done, the writer asked how they knew the prover was sound, meaning the theorems it proves are actually true.
How it went
Claude checked and immediately found five different ways it could “prove” false theorems with the assistant, and fixed them. It then proposed a fuzzer to find more soundness problems, but on testing admitted the fuzzer it proposed would have been completely useless. The writer calls the whole experience “a constant fractal of hallucination”, a constant struggle to keep their own and Claude’s grip on reality.
Try it yourself with Claude
Here's a proof assistant I just built: [description]. Before I trust it, check whether the prover is actually sound; can you find any way to 'prove' a false theorem with it? Fix anything you find.
A starting prompt written by Agents in the Wild. Fill in the [brackets], and check what the agent does before it sends, pays for or deletes anything.
After $45 in GitHub Actions overage in one month, a solo founder put an idle homelab to work as CI runners, with Claude making it easy to set up. Daily spend dropped from $2-4 a day to under $0.25.
A first homemade synth has 36 hall-effect keys and buttons and 7 rotary encoders. The firmware started by hand, then gentle prodding at Claude got everything needed much faster. Rev 2 of the PCB is nearly done.
Working on a Quake project, Claude decided on its own to build a test suite: demo playbacks with snapshots compared pixel by pixel, plus recorded demos whose screenshots it checked for each element.
Four years after shelving a homemade RSS reader and selling off its $500 domain, the writer built it this year with Codex and Claude. They use it daily, along with 30 other strangers who found it public.
Chasing a childhood itch for the exact 1996 After Dark Flying Toasters screensaver, the writer turned to Claude. It did a lot of good reverse-engineering work, but kept inventing details that needed pushback.
A rabbitmq connector occasionally got stuck after rare network drops, too minor to spend days fixing. Claude built a test harness that reproduced it, and the actual fix wasn't too hard.