Core question
When numerical observations cover only local regions, can global knowledge landmarks support more stable generalization across the full space?
Project 03
knowledge_landmarks_toy
A small-scale implementation used to examine a softer form of structural prior, in which global knowledge landmarks may regularize sparse local observations.
Suggested code order
Read the trainer first, then inspect the landmark construction, and finally trace how the regularization term interacts with the model update.
Mechanism to watch
This setting is not about exact logical satisfaction. It can be viewed instead as a softer regularization mechanism driven by global structural knowledge.