“Supervised fine-tuning a model on data that does not game the environment for easily detectable behaviors like sycophancy can reduce the likelihood of reward tampering in holdout environments.”