autoresearch

To automate ML experiments overnight, you can use an autonomous loop that iteratively patches train.py, executes fixed 5-minute GPU runs, and logs results. This enables continuous architecture exploration and hyperparameter tuning while you sleep.

Git ratcheting for reproducible machine learning is a workflow that commits only experimental improvements to train.py. It preserves a monotonic history of successful changes, ensuring traceability and reproducibility of your autonomous ML research results.

To set up an autonomous experiment loop, you need a locked evaluation harness in prepare.py and precise directives in program.md. Running the loop then patches train.py, executes time-budgeted experiments, and logs outcomes to a persistent results.tsv file.

Yes, autonomous architecture exploration can run with a strict time budget by executing fixed 5-minute GPU experiments. The AI-driven loop patches your training script, evaluates the modification, and only commits changes that yield improvements.

An autonomous ML agent decides which training script changes to keep by evaluating each 5-minute GPU run against a locked evaluation harness. It uses git ratcheting to commit only modifications that produce measurable improvements, discarding unsuccessful patches.

Running autonomous ML experimentation on a single-GPU machine limits you to sequential, fixed 5-minute experiments rather than parallel runs. It is designed for overnight research and requires a locked evaluation harness to ensure valid, reproducible comparisons across iterations.