Get The Proof Receipts

The companion to the video “Claude Stopped Testing Code. It Started Proving It.” INSIDE: (1) TEST vs PROOF in one picture: why a fuzzer that tries 1.5 billion inputs still can’t promise anything about the rest, and what a machine-checked proof covers instead. (2) THE NUMBERS from Anthropic and Theorem’s Oct 22, 2026 research: 239 entry points across 123 libraries proved crash-free on every input under 64 MB, ~180 million lines of Rocq proof, seL4’s 2 vs 20 person-years, Opus 5.5 using ~40% fewer tokens than Opus 5, and 17% fewer when agents share proofs. (3) THE BUGS proving found that fuzzing missed, from a ~20 MB-only decompressor bug to a rediscovered libexpat CVE. (4) THE CATCH: what “formally verified” does not mean, and the three things you still have to trust. (5) OUR OWN TEST, step by step: a fuzzer threw 10 million inputs with 0 crashes, then Claude Code and the free CBMC checker found a ~63 MB counterexample, fixed it in one line and proved the fix, with the exact commands to reproduce it. (6) Every link from the article: the published proofs, the RISC-V and Linux semantics, Theorem’s pipeline and the Claude skill.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — three practical AI emails a week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.