[bibtex key=EnayatHamkins2018:Ord-is-not-definably-weakly-compact]

In ZFC the class of all ordinals is very like a large cardinal.  Being closed under exponentiation, for example, Ord is a strong limit.  Indeed, it is a beth fixed point. And Ord is regular with respect to definable classes by the replacement axiom.  In this sense, ZFC therefore proves that Ord is definably inaccessible.  Which other large cardinal properties are exhibited by Ord? Perhaps you wouldn’t find it unreasonable for Ord to exhibit, at least consistently with ZFC, the definable proper class analogues of other much stronger large cardinal properties?

Meanwhile, the main results of this paper, joint between myself and Ali Enayat, show that such an expectation would be misplaced, even for comparatively small large cardinal properties. Specifically, in a result that surprised me, it turns out that the class of ordinals NEVER exhibits the definable proper class analogue of weak compactness in any model of ZFC.

Theorem. The class of ordinals is not definably weakly compact. In every model of ZFC:

1. The definable tree property fails; there is a definable Ord-tree with no definable cofinal branch.
2. The definable partition property fails; there is a definable 2-coloring of a definable proper class, with no homogeneous definable proper subclass.
3. The definable compactness property fails for $\mathcal{L}_{\mathrm{Ord,\omega}}$; there is a definable theory in this logic, all of whose set-sized subtheories are satisfiable, but the whole theory has no definable class model.

The proof uses methods from the model theory of set theory, including especially the fact that no model of ZFC has a conservative $\Sigma_3$-elementary end-extension.

Theorem. The definable $\Diamond _{\mathrm{Ord}}$ principle holds in a model of ZFC if and only if the model has a definable well-ordering.

We close the paper by proving that the theory of the spartan models of Gödel-Bernays set theory GB — those equipped with only their definable classes — is $\Pi^1_1$-complete.

Theorem. The set of sentences true in all spartan models of GB is $\Pi_{1}^{1}$-complete.