Reward hacking grows with model scale; a sound certify-or-abstain gate blocks it by construction (three pre-registered GRPO studies)
Daniel Culotta, independent researcher. Every number below comes from scripts on disk and two adversarial audit passes that re-ran them. The "Where it fails" section is not a disclaimer at the end. It is the point. Written with LLM assistance: the experiments, the pre-registrations, the analysis and every number are...
Aug 231