Alignment is not an end state specifiable ex ante but an iterative process requiring constant refinement, feedback, and learning — and the coordination layer is where this process actually lives. Unlike model-layer alignment where a single implementation error in a monolithic system can be catastrophic, coordination-layer modularity means single-point failures are bounded by design. The risks of reward hacking are far too high if alignment is mischaracterized as a one-time training-time destination rather than a continuous runtime process.