It's surprising how little interest there was in qualitatively inspecting gpt-oss CoTs. I mean these are they guys who created the paradigm, I guess they're not using GRPO variants like ≈everyone else, are there differences? Nope, people only care about capabilities.