Debiasing Reward Models by Representation Learning with Guarantees | Not Hacker News!