Core question
How can a valid output structure become a differentiable objective without introducing a teacher model?
Project 02
semantic_loss_toy
A small-scale implementation used to examine how logical validity can be expressed directly as probability mass over satisfying worlds.
Suggested code order
Begin with the constraint definition, then inspect the loss computation, and finally trace how the loss enters the trainer.
Mechanism to watch
A key mechanism is that optimization is not over the most likely class alone; it is over the total probability mass assigned to worlds that satisfy the logical constraint.