BYOC Protocol (Bring Your Own Code)

To integrate your existing training code with AgentCommander, follow this contract.

1. Interface

Your strategy.py must save weights and be loadable by the evaluator.

# strategy.py
def load_trained_model(path, device):
    model = MyModel()
    model.load_state_dict(torch.load(path))
    return model

2. Evaluation

Your evaluator.py must print exactly: Best metric: 0.123 (Lower case 'm').

3. Data Safety

Do not modify test data in memory. The system includes anti-cheating checks that will invalidate your score if you do.