Gating Claude's Own Rewrites

Video 3 moves the same proof gate from a human committing code to an AI writing it, wiring FMT into Claude through an MCP connector and a standing rule in the project's CLAUDE.md: nothing Claude rewrites counts as done until it's proven equivalent to what it replaced. Across three independent runs, Claude tries a rewrite that looks right and isn't (FMT catches it in seconds), a completely different rewrite that's correct on the first try, and — most tellingly — two further rewrites that come back "equivalent" at only 57% confidence, which the gate treats the same as a failure until Claude tries again and gets a rewrite proven outright, confidence 1.0. The point isn't that FMT picks the prettiest code; it's that nothing ships until it's shown, not guessed, to behave exactly like the code it replaced.