We know offline training -- pretraining, dpo… data is clear in advance. We also know online training -- ppo, grpo... data is built while training. New: Humanline training -- any data (offline/online) shaped to match human perception → can yield online perf at lower cost