Did you not read the article? IIUC the entire chip is formally verified (because, why not)
Formally - but by LLM. In theory practice is the same as in theory.
Will be cool to see if Linux can run on a QEMU or FPGA based implementation of this and how the efficiency numbers look then because it's not quite as interesting if it's not general purpose computing ready.
Does the proof correspond with the chip? Does it prove useful properties? Are the properties it proves desirable? Desirable for what? How did you check?