We started with full RLHF, found the infra burden heavy for our team size, and DPO got us 80% of the alignment win with a fraction of the complexity. RLAIF is tempting for the data flywheel. What post-training approach did you ship to production, and what would you choose if you started over?