The world increasingly trusts AI to write code. But if the author is a machine you can't trust on its word, how do you know what the code actually does? Today's answer is "read it and hope." I wanted better.
LOOM is a tiny language that makes what code is allowed to do machine-checkable, and refuses to run anything that lies about it. "AI proposes, the compiler disposes."
In a few days it grew into a trust layer for AI-written code: it checks not just what code does, but who vouches for it — demanding independent, non-AI corroboration. Open (MIT), self-verified by 221 checks that can only grow greener, and now cross-platform — one verified program runs identically on Python and JavaScript.
The unusual part: I didn't grow it alone. An autonomous research organism develops it daily — and this week it learned to pre-check its own proposals in a sandbox and adversarially attack them before I even look. It proves its work; it doesn't ask me to trust it. How the engine works, I'm keeping to myself.
Built solo, in the open, from Ukraine 🇺🇦 — ⭐ github.com/umbraaeternaa/loom
Top comments (0)