Reasoning levels are often just injected system prompts so not a great way to fingerprint models.