Run Train

When you've already figured out your training command and just need to run it with proper evidence collection, this is your tool. It handles startup checks, short verification runs, full training kickoffs, and resumes, then writes everything (commands, logs, configs, checkpoints, metrics) into a standardized train_outputs directory. It's not for environment setup or exploratory sweeps. The boundaries are refreshingly clear: you bring a runnable command and a goal, it executes conservatively and gives you structured output including status.json and a scientific changelog. Built for deep learning repos where reproducibility context matters more than moving fast and breaking things.