Machine learning models often pick up hidden shortcuts: a résumé screener may use gender, or a chatbot may produce harmful text, despite the stated goal. Rules added to the training objective are hard to enforce inside a neural network.
Mechanistic Shaping instead states rules through the model's internal activity patterns. During training, it penalizes the model when those patterns express an unwanted concept, or when the concept drives a decision. The method was tried on image, medical, and language tasks with several training styles.
Targeted behaviors dropped while task performance held. The unwanted concept usually stayed detectable after retraining the detector on the new representations, so the rule can be verified.