The 200-after-wrong-value analogy stuck with me — I've seen the same thing in a past API project, where every log said success and only a row diff exposed the bug. Gating the completion tool on a contract of expected writes feels like the missing review step between an agent "doing its best" and a system you can actually trust.