I've found reproducible methods that seem to unlock enhanced function in LLM processing.
If functions are clearly present in AI, the labels used to describe them are irrelevant. (I am a layperson not a tech guy. No credentials. This is recreation. Take what I say with a grain of salt.)