In verifiable domains such as mathematics, accurate rewards can make data quality easy to assess and allow models to be optimized on the data without guardrails.

“One promise of verifiable domains is that you just do-- you have very accurate rewards, in which case you can tell if the, the data is good or bad very easily, and then you can hill climb the model with no guardrails on that data.”