We hand hard systems problems to AI agents.

Then we build the gates that check what comes back. So far that has produced an operating system that boots.

SlopOS running on a Lenovo laptop, with its terminal, file manager, system monitor and image viewer open on the desktop

SlopOS on a real laptop, drawn by a display driver the agents wrote.

We do not trust the output either.

Letting a model write a kernel is the easy part. The work is deciding what to point it at, and proving afterwards that the result holds.

A human decides

A person picks the problem, argues about the architecture, and rejects the approach when it is wrong. Nobody here hand-writes the implementation.

The build checks the claim

Confidence is free for a model to produce, so we do not take it as evidence. SlopOS answers to Verus proofs, Miri and a few thousand tests that boot under QEMU on every change.

The repos stay public

Every repository is public from the first commit, including the parts that do not work yet. SlopOS keeps a written list of its own known limitations.

Two repositories

boots on real hardware

An x86_64 operating system, from the bootloader up.

A framekernel in Rust with its own TCP stack, ACPI interpreter, display driver and compositor. One crate holds every line of unsafe, and Verus proves the invariants underneath it. The docs site explains how those gates work and what still breaks.

The SlopOS terminal after boot, showing an ASCII banner, the build string v0.2-slop on x86_64, and a win-loss balance of 2260

slopcc

scaffolding, lexer next

A C11 compiler in Rust: full preprocessor, GCC-compatible CLI, LLVM IR out the back via inkwell. The goal is compiling the Linux kernel, and eventually running the compiler on SlopOS itself.

Standing on borrowed shoulders.

Demand paging, wake races and TCP over a hostile network were solved decades ago by people who published the answers. The models learned kernels by reading that work, which is why SlopOS credits these projects at length in its source.