And in DEVELOPER_POLICY: "思考過程や for thinking blocks" but it also says the format requirement depends on the user/task, and thinking_mode being enabled means the model "must always output a thinking block".
DEVELOPER_POLICY explicitly says don't output think tags and don't output thinking.
These two ARE in conflict. So per the priority rules, ROOT_SYSTEM_POLICY wins.
OK so I should output thinking. Let me do that.
Actually, let me look at the actual user instruction: "