Formal Verification is Not Verifiable · Alperen Keles · personal blog, November 13, 2026
"Formally verified" should come with a list of what was proven
Keles writes this as a follow-up to his December 2025 post "Test, don't (just) verify". His argument is that testing is easy to check and formal verification is not. A bug report is an object a person can review. A proof is checked by machine, but its definitions, the models of the system it reasons about and the prover itself still have to be trusted, and "formal verification makes forall claims, testing makes exists claims", so one mistake in a trusted part can sink the whole result. Binary-level verification helps by removing the source-language semantics and the compiler from the trusted base, he writes, but machine semantics and Linux process semantics stay in it. His proposal is to stop saying "formally verifying X" and say "formally verifying X properties of Y". He is still optimistic. He now believes single-threaded program verification can be solved, with binary-level proofs for any first-order, universally quantified property that can be written as a property-based test. He expects multi-threaded and distributed systems, higher-order and non-functional properties, and real deployments to take years more work.
Why it matters: If an agent hands you a proof, the proof is the part you do not need to review; the specification and the trusted base are. Keles's wording, "X properties of Y", is a usable acceptance rule for verification gates: a result that cannot say which properties it proved, and what it assumed, has not told you much.