Most errors are probably responses that didn’t finish before their 3K token limit. They’ve measured how well RL is able to shorten the response to their limit.