Jorge RoldanMechanistic interpretability and AI safety.https://jorgeroldan.io/What synthetic-document fine-tuning transmits, and what it doesn't: a layered dissociation in evaluation-aware behavior on Qwen2.5-7Bhttps://jorgeroldan.io/research/sdf-dissociation/https://jorgeroldan.io/research/sdf-dissociation/Synthetic-document fine-tuning on Qwen2.5-7B installs the corpus belief at the recognition layer, not in generation, and the only task-behavior effect attaches to the wrong cue.Tue, 19 May 2026 00:00:00 GMT