Is it confirmed that synthetic data was used for gpt-oss training? I didn't pick up on that in the press release or see it elsewhere. Did I miss it or is Sean speculating that it is the case?