Autoresearch

This is an autonomous ML research loop that runs GPU experiments while you sleep. Point it at a train.py file, write research directives in markdown, and it runs 5-minute experiments, evaluates on a single metric, and uses git ratcheting to only commit improvements. You wake up to 100+ logged experiments and a monotonically better model. The workflow is opinionated: fixed time budgets, single metric optimization, automatic commits. It's designed for the Karpathy school of thought where you spend less time writing Python and more time writing directions. Best suited for when you have spare GPU cycles overnight and want to explore hyperparameter or architecture variants without babysitting runs. The ratcheting mechanism means you can't regress, which is either exactly what you want or too constraining depending on your research style.