The Darwin Gödel Machine lets a coding agent modify its own codebase and judges each proposal by benchmark execution rather than a proof of improvement. Viable variants are retained in an archive, so search can branch from prior discoveries, and the best archived score improved over the course of the search. Compositional Intelligence draws the architectural lesson cautiously: open-ended exploration needs evaluators and lineage, but no single branch is guaranteed to improve monotonically.
Schmidhuber's Gödel machine rewrites itself only when it can prove the rewrite is an improvement. Proof is the part that has never been practical.
This system drops it. A coding agent edits its own codebase, and each edit is judged by running it on benchmarks. Every variant is kept in an archive, including the ones that lost, so the search can branch from them later.
Scores on SWE-bench and Polyglot climb over successive self-modifications. The archive is doing the work an open-ended search needs: keeping the material that looks useless now.