AutoResearch runs on files:
prepare.py: the fixed harness. Contains constants, the evaluation function, dataset loading, and the tokenizer. You never modify this file.train.py: the sole editable file. Contains the model architecture, optimizer, hyperparameters, and training loop. This is what your agent changes.program.md: your research instructions in Markdown. Defines the objective, constraints, and strategies for your agent.
Each file has a specific boundary. prepare.py sets the rules. train.py is the playground. program.md is your voice.