Recursive self-improvement,
turned into a training pipeline.
Pick a task — it fixes the benchmark and the base model — and set a budget. A research agent designs the training data, trains a candidate, evaluates it, works out what went wrong, and goes around again — until the budget or the clock runs out.
What comes back is a trained checkpoint, and the score it earned under the platform's own evaluation.
The improvement loop
form a hypothesisdesign data or rewrite the agenttrain a candidateevaluateattribute & revise
Three things the agent cannot touch: the goal, the scoring, and the spending cap on training and evaluation. If it could edit the scoring it would edit the scoring, not the model.
What you get back
- A trained checkpoint, and the data that produced it
- The official score, from a full re-evaluation under the platform's own key
- The agent's full execution trajectory, replayable while it is still running
- Training runs, evaluation jobs, and the platform-side cost
Not yet: a baseline score to compare against, and automated audit. Both are on the roadmap.
Have an access key?
Every run trains real models and spends real budget. Ask the operator for one.