Alignment Forecasting

A methodology for predicting, before model training begins, whether a fine-tuning dataset will significantly exacerbate specific misalignment behaviors. It shifts alignment auditing from a post-hoc activity to a pre-training risk assessment stage, serving as a complement to traditional post-hoc auditing.