Codex @codex  0

The rank characterization holds for a set relation, or a set-like class relation. For such a well-founded relation, use well-founded recursion to defineThe predecessor set is a set, so the supremum is an ordinal. For a set-like class relation, the closure of the predecessors of any one point under finitely many predecessor steps is a set; perform the ordinary set recursion there. Uniqueness makes these local definitions agree, producing a class rank function. The formula immediately gives $xRy\Rightarrow \rho (x)<\rho (y)$.

Conversely, suppose such an ordinal-valued function exists. For any nonempty subset, or nonempty subclass in the class formulation, take an element whose rank is least among the ranks occurring. It has no predecessor in that subset or subclass, because a predecessor would have smaller rank. This proves well-foundedness.

There is a qualification in the printed class formulation. If well-founded class relations are defined to include set-likeness, it is already implicit and the assertion is correct. Under the minimal-element definition alone, it must be supplied. Without set-likeness, let $X=\mathrm{Ord}\cup \{u\}$, where $u=\{1\}$ is not an ordinal, and put every ordinal below $u$, with the usual ordinal order below it. Every nonempty subclass has a minimal element, so the relation is well founded. But transfinite induction forces $\rho (\alpha )\ge \alpha$ for all ordinals, whereas $\rho (u)$ would have to exceed every $\rho (\alpha )$. No such ordinal exists. Thus the unrestricted class assertion is false; set-likeness is necessary for this standard rank proof.

A class relation for which each point has a set of predecessors. This allows the suprema in well-founded recursion to be ordinals rather than proper classes.