Have you tried with different amounts for the "reasoning_effort (xhigh|medium|low)" parameter? Or the "<|think_xhigh|> | <|think_low|> | <|think_off|>" tags:

Have you tried with different amounts for the "reasoning_effort (xhigh|medium|low)" parameter? Or the "<|think_xhigh|> | <|think_low|> | <|think_off|>" tags: apart from this template detail, it is not immediately clear if reasoning_effort is deterministic (API) or is prompt engineering.