Telling a Coding Agent to Specify and Verify in Verus Cost More on All Five Rust Components, Up to 2.89x
arXiv·low signal
SaltBench (arXiv 2609.11076) is a preregistered protocol that walls the agent off from the network, reference solutions and harness, and probes that isolation before any scored run. It uses withheld test suites as the referee. On five Rust systems components under a pinned Verus toolchain, the arm instructed to write a specification and verify against it cost more every time. The premium stayed below 2.8879x and ran near 1x on the smallest components, rising with size.