How Ari runs autoresearch
one research cycle per experiment
Replay
Pause
Cycle 1 of 5
1
Run experiment
→
2
Pull traces
→
3
Launch judges
→
4
Update notebook
→
5
Revise hypotheses
↺
Experiment
Eval traces
1,600 rollouts, 8 per square
Notebook
kept by Ari
Hypotheses
Experiment log
Ari Autoresearch on Lichess puzzles. Failure-cause shares are illustrative.