We're a research company first. These are our results on public benchmarks.
View Benchmarks →Score on validation set shown above. For the holdout test sets, PhantomWiki progressed 41% → 65%; HoVer, 56.9 → 85.8. See our methods below.
Each benchmark starts from a third-party program written by researchers — not a strawman baseline. We extract the benchmark dataset and split it into training, validation, and test sets. The test set is touched exactly once, at the very end.
The engine runs with an instruction set combining optimization constraints and seed ideas (architecture changes, retrieval strategies). GPT-5.4 nano is held constant throughout — every point of improvement comes from prompt and code changes, not a bigger model.
CodeEvolver combines an evolutionary optimizer with a reflection agent that analyzes failures and a coding agent that executes changes. Maintaining a diverse pool of candidates avoids the greedy optimizer problem, where improving a good program gets stuck before reaching a great one.
Every accepted change is preserved with its full experiment history: what was proposed, what it scored, and what it replaced.
Our research paper is in progress.