What Cross-Entropy Is and Is Not

The finale states the register split: exact one-hot selection and rational probabilities, with the logarithm named. It is one loss, not a broader claim.

highlighted = computed this step

What is exact

The exact register contains the one-hot target, the rational probability vector, the sum-to-one check, and the structural selection of the correct class probability 1/2.

p_k=1/2\quad\text{selected exactly}

What is named

The loss value is named as -log(1/2). The render pins no logarithm decimal. The required note states exact one-hot selection plus named log boundary.

H=-\log(1/2)\quad\text{named}

This is one loss on one pinned probability distribution. It is exact up to the named log boundary; it is NOT learning, NOT a softmax computation, and NOT a broader model claim.

\text{one exact-or-named loss, deferred claims explicit}