RLHF / RL
RLHF / RL trains the model with human or rule feedback; badly designed feedback can teach reward hacking, so evals still matter.
When you will meet it
Badly designed feedback teaches reward hacking — one more reason Stage 7 insists on evals.