I read that an AI model proved 239 entry points in 123 open-source libraries can never crash, for any input under 64 MB. Not tested. Proven. Then, under limitations, the line that stopped me: the proofs are about single-threaded, static RISC-V binaries linked against musl, while most of us run them on x86 or ARM with glibc.
I like this more than I expected. It is the deterministic gate taken all the way: 180 million lines of proof no human reads, and a small checker that verifies every one without a model in the loop. What you trust shrinks to three things: the machine model, the theorem and the checker. All three are public. But a proof holds on the machine it describes, not the one in our racks. So when a vendor says verified, I ask which chip, which libc, which theorem. No answer means they tested it.