You're giving your model instructions that it's literally incapable of understanding. A random word selection lottery machine will never do anything meaningful to determine if a source is primary or secondary.

It depends on the odds of the lottery. As a straight-up classification task I'd expect it to do better than chance, which might not be good enough for you.

It is admittedly awkward to bend to the machine to get what you want. I see these kinds of constraints like given above as an impetus to push the LLM designers to rise to the occasion, assuming they’re listening to all the prompts funneling back their way. This may all be wishful thinking however but hopefully someday these kinds of prompt constraints will be satisfied.